Channel encoder apparatus and channel decoder apparatus

Abstract

There is disclosed a channel decoder apparatus (40) for generating a vectorial feature representation as a vectorial feature representation (40a, 45) in a format specific of a source decoder (50), from a sequence of symbols (36), the vectorial feature representation (40a) representing a source signal (1, 5), the channel decoder apparatus being configured to apply at least one channel-decoder neural network, NN (40a, 42), to the sequence of symbols (36) or another representation (46a, 43a, 42a) of the sequence of symbols, thereby generating the vectorial feature representation in the format specific of the source decoder (50). There is disclosed a channel encoder apparatus (30) for encoding a source signal (1) received as a vectorial feature representation (20a, 21a) from a source encoder (20), the vectorial feature representation (20a, 21a) being in a vectorial format specific of the source encoder (20), the channel encoder apparatus (30) being configured to apply at least one channel-encoder neural network, NN (30a), to the vectorial feature representation (20a, 21a) in the vectorial format specific of the source encoder (20) or to another feature representation (22a, 23a, 24a, 25) of the vectorial feature representation (20a, 21a), to provide a sequence of symbols (35) from the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25).

Description

[0001] FHIIS24EM36-2024373152. DOCX 1

[0002] Channel encoder apparatus and channel decoder apparatus

[0003] The present invention refers to techniques regarding the channel encoder and decoder apparatuses, and encoder and decoders. Some of these techniques are called here SplitJSCC (see below).

[0004] Examples of channel decoding techniques are provided in Figs. 6, 7, and 8 according to the prior art. A detailed description of the prior art is provided below. However, it is already noted that Fig. 6 shows an example of a source encoder 620 and a channel encoder 630 which is not implemented with a neural network.

[0005] The same applies to the example of Fig. 8. Fig. 7 shows a theoretical example of the prior art according to which there is no separation between a source encoder and channel encoder, and source decoder and channel decoder.

[0006] It is difficult to use a universal channel encoder, because, if the input to the channel encoder is a sequence of bits, errors (or reduction of precision) may be caused. A channel encoder 630 as in Fig. 6 provides a suboptimal result, since there is no semantic relationship between the source signal and the sequence of bits that are compressed.

[0007] It would be preferably to find out strategies for applying a channel encoding and channel decoding, which are more responsive to the source signal.

[0008] This issue is exacerbated by the fact that the source encoders are not univocal, but they can vary based on the particular application. Different source encoders (or the same source encoder in different operational modes) may therefore provide values whose statistical properties are pretty different from each other. The statistical properties of the sequences or bits generated by the different source encoders could be not mirrored by the channel encoder and, therefore, a waste of information could arise.

[0009] There are some non-NN based channel encoders I decoders, which are however agnostic of the source encoders / decoders, and therefore are not optimized for the particular type of signal and for the particular source encoder / decoder used.

[0010] A solution to these problems, or at least a reduction of their inconvenience, is pursued. FHIIS24EM36-2024373152. DOCX 2

[0011] Summary

[0012] In examples there is provided, inter alia, a channel decoder apparatus for generating a vectorial feature representation as a vectorial feature representation in a format specific of a source decoder , from a sequence of symbols, the vectorial feature representation representing a source signal , the channel decoder apparatus being configured to apply at least one channeldecoder neural network, NN , to the sequence of symbols or another representation of the sequence of symbols, thereby generating the vectorial feature representation in the format specific of the source decoder .

[0013] In examples there is provided, inter alia, a decoder configured to change among a plurality of source decoders, to thereby select a channel-decoder NN based on the selected source decoder.

[0014] In examples there is provided, inter alia, a channel encoder apparatus for encoding a source signal received as a vectorial feature representation from a source encoder , the vectorial feature representation being in a vectorial format specific of the source encoder the channel encoder apparatus being configured to apply at least one channelencoder neural network, NN , to the vectorial feature representation in the vectorial format specific of the source encoder or to another feature representation of the vectorial feature representation , to provide a sequence of symbols from the vectorial feature representation or the other feature representation .

[0015] In examples there is provided, inter alia, a user equipment, UE, comprising the channel encoder apparatus and the source encoder apparatus, wherein the channel encoder apparatus is configured to transmit the sequence of symbols from the vectorial feature representation or the other feature representation in the uplink, the UE being configured to send signalling indicating that it has the channel encoder apparatus with the channel-encoder NN . FHIIS24EM36-2024373152. DOCX 3

[0016] In examples there is provided, inter alia, a user equipment, UE, comprising the channel decoder apparatus and the source decoder apparatus, wherein the channel decoder apparatus is configured to generate a vectorial feature representation , as a vectorial feature representation specific of a source decoder, from a sequence of symbols , the vectorial representation representing a source signal, in the downlink, the UE being configured to send signalling indicating that it has the channel decoder apparatus with the channel-decoder NN .

[0017] In examples there is provided, inter alia, a radio access network, RAN , comprising the channel encoder apparatus and the source encoder, wherein the channel encoder apparatus is configured to transmit to a user equipment, UE, the sequence of symbols from the vectorial feature representation or the other feature representation , the RAN being configured to send signalling requesting the UE to send information whether the UE has the channel encoder apparatus with the channel-encoder NN or whether the UE has the has the channel decoder apparatus with the channel-decoder NN

[0018] In examples there is provided, inter alia, a encoder comprising the encoder of any of claims 31-53, comprising the channel encoder apparatus and the source encoder.

[0019] In examples there is provided, inter alia, a method for generating a vectorial feature representation as a vectorial feature representation in a format specific of a source decoder , from a sequence of symbols , the vectorial feature representation representing a source signal , the method including applying at least one channel-decoder neural network, NN , to the sequence of symbols or another representation of the sequence of symbols, thereby generating the vectorial feature representation in the format specific of the source decoder

[0020] In examples there is provided, inter alia, a method for encoding a source signal received as a vectorial feature representation from a source encoder , the vectorial feature representation being in a vectorial format specific of the source encoder , FHIIS24EM36-2024373152. DOCX 4 the method including applying at least one channel-encoder neural network, NN , to the vectorial feature representation in the vectorial format specific of the source encoder or to another feature representation of the vectorial feature representation , to provide a sequence of symbols from the vectorial feature representation or the other feature representation .

[0021] In examples there is provided, inter alia a non-transitory storage unit storing instructions which, when executed by a processor, cause the processor to perform a method..

[0022] Fig. 1 shows an example according to the invention.

[0023] Fig. 2 shows a variable according to the invention.

[0024] Figs. 3A, 3A and 3C shows examples according to the invention.

[0025] Figs. 4A shows an example of neural network according to the prior art.

[0026] Figs. 4B and 4C show examples of neural networks according to the invention,

[0027] Fig. 5 shows an example of communication according to an invention,

[0028] Figs. 6, 7, and 8 show examples according to the prior art.

[0029] Fig. 9 shows an example according to the invention.

[0030] Figs. 10A-10D show tables according to the invention

[0031] Figs. 11 and 12 show models for comparing the invention and the prior art

[0032] Figs. 13-15 show results of comparison between the invention and the prior art.

[0033] Here below, in Figs. 11 and 12, reference numerals 351 and 361 are equivalent baseband representations of the continuous transmitted (or received, respectively) signal.

[0034] Here below examples are provided according to which a selection is performed of the channel-encoder neural network based on the source encoder.

[0035] Fig. 1 shows a general example for showing the present technique. An encoder 2 may be a media encoder (e.g., speech, audio and / or video encoder) which provides an output bitstream e.g., as wireless transmission (e.g. wireless terrestrial transmission or wireless satellite transmission, e.g. satellite-to-satellite transmission or satellite-to-earth transmission or earth-to-satellite transmission). The bitstream may be the form of a FHIIS24EM36-2024373152. DOCX 5 sequence of symbols 35 (and in some cases also in symbols 35b, see also below) (Fig. 1 does not consider a pulse shaping filter 350 which may be notwithstanding appreciated in Fig. 11, which may be inputted with the symbols 35 to provide the pulse-shaped versions 361 of the symbols 35). In general, the symbols 35 (351) may be understood as modulation symbols for the wireless transmission. Notwithstanding, it is also possible to generalize the invention by taking into account more general symbols (e.g., bits or other representations, e.g. like real or complex numbers). The symbols may include, for example, pulse amplitude modulation (PAM) symbols or other kinds of symbols. The input signal is here called source signal 1. The source signal 1 can be, for example, a media signal, such as a video and / or audio signal. In the case of an audio signal, the source signal 1 may provide information on the audio signal for, for example, a time frame of 20 ms (or different time length). In the case of video... in the case of text... The encoder 2 may include a source encoder 20. The source encoder may apply a source encoder neural network (NN) to the source signal 1. The output of the source encoder 20 may be a vectorial feature representation 20a. The vectorial feature representation may be a vectorial latent representation. Downstream to the source encoder 20, there is provided a channel encoder apparatus 30 (inputted with 25, which may be a processed representation of the representation 20a). The channel encoder apparatus 30 may provide the output 35 (361), e.g., as a sequence of symbols. The channel encoder neural network may include at least one NN. As it will be explained later, the neural network may constitute the main element of the source-matched channel encoder apparatus 30. This is because there may be many instantiations of the neural network used by the channel encoder 30 and the neural network may be associated to the particular source encoder 20 which is used, so that a source-encoder NN instantiation is chosen among a plurality of NN instantiations. In several examples, therefore, for any source encoder 20 which is selectively chosen for the encoder 2, there will be an appropriate source-matched channel encoder NN, matched to the source encoder. The output 35 (361) (sequence of symbols) of the encoder 2 (and of the source-matched channel encoder 30) may be transmitted as wireless signals towards a receiver. The receiver may be a decoder 4. The input signal 36 (361) may be affected by noise 35', which is unwantedly added, at the unwanted adder 35aa, to the sequence of symbols 35 (351), to have a modified version 36 (361) of the sequence or symbols 35 (351). The decoder 4 may therefore be inputted with the input sequence of noisy symbols 36 (361) and may apply a channel decoder 40. The channel decoder 40 may have at least one neural network applied to the input sequence of symbols 36 (or in their pulsed shaped version 361; Fig. 1 does not show the inverse FHIIS24EM36-2024373152. DOCX 6 pulse shaping filter 360, which may be notwithstanding appreciated in Fig. 12). The output 45 of the source-matched channel decoder 40 may be inputted into a source decoder 50. The source decoder 50 may have at least one neural network applied therein. The output 5 of the source decoder 50 may be the output of the decoder 4 and may be understood as a decoded audio or media or video signal rendered or to be rendered. The source decoder 50 may be interchanged by the decoder 4, e.g. in a similar way for which the source encoder may be interchanged by the encoder 2. Therefore, there may be different instantiations of the channel decoder 40, each for the particular instantiation of the source decoder 50.

[0036] Both the output 20a of the source encoder 20 and the input 45 of the source decoder 50 may have a specific format associated, respectively, to the source encoder 20 and the source decoder 50. The vectorial feature representation in the vectorial format specific of the source encoder or decoder is extremely important for the subsequent passages of the description. The source encoder 20 produces an output 20a which has a predefined format proper of the source encoder 20 and / or of the specific type of source signal 1 or, in some cases, of the particular mode at which a source encoder 20 operates. In the cases in which the source encoder 20 can output different vectorial formats, then these vectorial formats are normally standardized (e.g., with a particular range, in particular granularity, in particular format, and so on). The channel encoder 30 may be ingested with the complete vectorial representation 20a outputted by the source encoder 20, and the channel decoder 40 outputs the vectorial representation 45 e.g. in one single process. Said in other terms, the input 25 of the channel encoder 30 may be unsegmented among bytes (or words, or similar), but all the vector elements may be applied to the neural network in parallel, simultaneously.

[0037] (Fig. 2 shows an example in which further data lb, such as non-media data which notwithstanding may pertain to the signal la, such as side information, metadata, etc. may be provided to a non-source-matched channel encoder 30b, e.g. non-based on NNs). The further data lb may be encoded as signals 35b (read by the decoder 4 as signals 36b). The signals 36b may be provided to a non-source-matched channel decoder 40b and provided as further data 5b (l'b). The further data lb (5b) may include information on the signal la which is not sematic, and may include, for example, information on the length of the signal la, header information, etc. FHIIS24EM36-2024373152. DOCX 7

[0038] An example is provided by comparing Figs. 4A and 4B. Fig. 4A shows a channel encoder neural network 30PA according to the prior art. Here, the training has been performed without agnostically of the source decoder of the type of source signal, i.e. independently of the source encoder and of the type of source signal. For example, several bytes are sequentially ingested to the neural network 30PA. At the first iteration, a first byte is ingested; in a second iteration, a second byte is ingested and so on. There is no semantical meaning in how the NN 30PA applies the weights to the bytes: it is only a separation of the vector elements onto strings to be encoded. Notably, every sequence can be in general independent from the other ones. Therefore, the weights in the layers of the NN are applied to each byte in a sequence, e.g. without memory of the past sequence. In the case of having three vector elements (vector components 1, 2, 3), each encoded in three bytes, we have nine iterations of the neural network 30PA.

[0039] On the other side, the inventive channel encoder 30 (in its NN 30a, 22, see also below) takes all the vector elements in their entirety and provides them to determined positions in the neural network 30a. In this case, as shown in Fig. 4B, there is only one single iteration in which all the vector elements are ingested in parallel, simultaneous inputs, and in one single iteration. The same applies (but in reverse direction) in the source-matched neural network (42, 40a) of the channel decoder 40: here, the input are symbols 36 (or representations of the symbols, may be in vectorial format) and the output (45 or 41a) is in the format of the input required by the source decoder 50. In this case, all the symbols 35 of a determined sequence are ingested to the neural network altogether, and the result may be valid for the order sequence.

[0040] The choice for this way of operating like in Fig. 4B would appear counterproductive: indeed, in Fig. 4A (prior art) we only use one single (e.g. universal) neural network 30PA, sequentially ingested by multiple bytes or words. Instead, in Fig. 4B, we need to have multiple neural networks 30a, 22 (multiple instantiations of the neural network), each adapted to the specific format of the source decoder. However, the inventors have understood that the prior art can introduce some errors due to the separation between bytes. The representation in bytes may be semantically incongruent with the nature of the source input 1 and, therefore, can generate errors. Indeed, the neural network in the prior art is trained on byte values which have nothing to share with the nature of the source signal 1. On the other side, with the example of Fig. 4B, the neural networks 32, 42, 30a, FHIIS24EM36-2024373152. DOCX 8

[0041] 40a, may be trained on the vectorial feature representation in the specific format that belongs to the considered source coding and, therefore, the result of the trained parameters (e.g., weights) may end having a semantical significance associated with the nature of the source signal 1. Notably, the neural networks 30a and 40a of the present technique may be trained using a backpropagation and a minimization of a loss function (e.g., distance loss 45v in Figs. 1 and 2) from a reconstructed version 45 (or 40a) of the vectorial feature representation (20a). The channel encoder apparatus 20 and the channel decoder apparatus 30 may be trained together, e.g. without optimizing the source encoder 10 and the source decoder 50. It shall be noted that it is both possible to have an offline backpropagation and minimization of the loss function, but in some other examples it is possible to have the backpropagation and minimization of the loss function during inference, as well, so that the source encoder NN 30a and the source decoder NN 40a are trained while operating. Instead, the source encoder and / or the source decoder may have constant parameters, and the channel encoder / decoder is optimized for it.

[0042] For the reasons above, the weights and the other parameters of the neural networks 32, 42, 30a, 40a have a semantical relationship with the vector elements, or said in other terms, their position is intimately related to the nature of the source signal 1 and of the specific format of the vectorial representation 20a. Therefore, the errors discussed above due to the subdivision in bytes and words does not impair anymore the inferences of the neural network.

[0043] The NN 30a, 42 may be fully connected, where the weights which are 0 are only obtained by training.

[0044] In some examples, in the example of fig. 4B, every node of every hidden layer is connectable to any node of any other hidden layer through a path, in backward or forward direction, constituted by non-zero weights (or weights which are 0 only as a result of training, but not as a result of a priori-choice).

[0045] In some examples, the NN 30a, 22 may have that nodes and weights in pre-defined positions specifically associated with the particular format of the source encoder / source decoder. In practice, the NN 30a, 22 is not a "general purpose" NN, but applies nodes and weights in association with the format. FHIIS24EM36-2024373152. DOCX 9

[0046] In other words, the succession of nodes and weights in one single hidden layer is in principle not repeated along the layer, but takes weights values which are only obtained by training for each node of each layer.

[0047] The examples of Fig. 4B are not necessarily exhaustive. For example, inventive solutions could also, like in Fig. 4C, include sequential networks (e.g. recurrent NNs). In a first iteration, only the first two vectorial components 1 and 2 may be inputted, and a status may be stored, e.g. for each layer. Then, in the second iteration, only a third vectorial component 3 may be inputted. In this second iteration, however, there is memory of the previous iterations, because the status (stored in the previous iteration) may have repercussions on the values in the layers. However, (keeping into mind the well-known property of recurrent NNs being generally developable onto connected NNs) if we develop the recurrent NN of Fig. 4C, we obtain the same of properties of the NN of Fig. 4B (which are here not repeated). Hence, the NN of Fig. 4C is, when developed, actually nonsequential.

[0048] In general terms, it may be understood that the NN is not separable into two independent sub NN for obtaining different portions of the output (which is symbols or other representation in Figs. 4B and 4C, but would be the representation of the symbols in the inverse NN at the decoder 40).

[0049] The NN 30a, 40a, 32, 42, may be understood as being optimized for the specific format of the signal outputted by the source decoder.

[0050] In some cases, we could call this inventive technical choice a choice of non-quantizing the vectorial components and / or the sequence of symbols at the input and / or output of the neural networks. It shall be understood, however, that this "non-quantizing" concept refers to the non-vectorial separation that could be done to the input and / or output of the neural networks. Of course, the processors implementing the neural networks will necessarily introduce quantization errors due to the very nature of processing binary values, for example. Notwithstanding, apart from the binary quantization (or more in general, processing quantization) impairing the processors, what remains is an unquantized application of the neural network. Further, the operations based on the neural network FHIIS24EM36-2024373152. DOCX 10 vectors may be considered lossless-compressed (apart, of course, the quantization error). It is understood that the output of the neural networks is in general terms a compressed version of the input, and this is due to the training.

[0051] Fig. 3B shows a first example of the example of Fig. 1. At first, input data 1 may include a source signal la. The source signal la may be inputted to the source decoder 20. As explained above, there may be several examples of source encoders or many instantiations or many operation modes. The output of the source encoder 20a may therefore be the vectorial feature representation in the vectorial format specific to the source encoder 20. It will be shown that the vectorial feature representation 20a can have several other representations (22a, 23a, 24a, 25) which are processed versions of the vectorial representation 20a and represent the source signal la. The versions 21a, 22a, 23a, 24a, and 25 are source-matched, since they are processed with at least one neural network, which are source-matched.

[0052] Downstream to the source-matched channel encoder NN 30a there is, for example, a physical transceiver to transmit the modulation symbols 35 (351) e.g. as electromagnetic waves or ultrasound. Dually, upstream to the source-matched channel decoder NN 40a there may be a physical transceiver to receive the modulation symbols 36 (e.g. in their version 361).

[0053] Upstream to the source-matched channel encoder neural network 30a, there may be other blocks (some neural network based and some deterministic) which are here discussed. Dually, downstream to the source-matched channel decoder NN 40a there may be may be other blocks (some neural network based and non NN-based) which are here discussed.

[0054] It possible to consider the at least one source-matched channel encoder NN as being instantiated by a cascade from by both the neural network 22 and the neural network 30a (in which case, we would have at least two neural networks), and in this case, the at least one first neural network 22 has, as input vectorial feature representation 20a. On the other side, it is possible to consider that the source-matched channel encoder NN is the only one indicated with 30a (and in that case, the value 25 would be the other vectorial representation of the feature representation 20a). In some cases, the training may be performed onto the signal 20a, while in other examples, in the training may be applied to FHIIS24EM36-2024373152. DOCX 11 the signal 25. The same applies for the source-matched channel decoder 40a and the source-matched decryption neural network 42 at the channel decoder apparatus 40.

[0055] With reference to the blocks in the channel encoder apparatus 30, a source-matched encryption neural network 22 (e.g. first channel matched channel encoder) may be applied to the vectorial feature representation 20a of the source signal la. The output of the source- matched encryption neural network 22 is indicated as the represented vectorial feature representation 22a. As explained above, the neural network 22 may be like the neural network of Fig. 4B. Downstream, a storage 23 may be provided. The output of the storage 23a may be inputted onto a TCP / IP block 24. The TCP / IP block 24 may be provided with further data lb. The further data lb may be non-media data which are provided to the TCP / IP for generating a TCP / IP standardized output feature 24a. A performance enhancing proxy 26 (PEP) may be applied to the standardized TCP / IP latent 24a to obtain a separated feature vector 25. Here, the output to the PEP 26 may be provided to two different channel encoders. At first, source matched channel encoder 30a (which substantially implements the channel encoder neural network, e.g. of Fig. 4B) which will encode the vectorial feature representation 45 onto a series of modulation symbols, while the non-media signal (e.g., the TCP / IP header 24) may be outputted and transmitted at side information, e.g. in association with the sequence of symbols from each SM channel encoder 30a. The symbols 35 (351) may be received (36, 361) by a decoder which is also a receiver.

[0056] The examples of neural networks shown in Figs. 4B and 4C are not restrictive. For example, the neural network 30a, 40a, 22, or 42 may be convolutional NN or may be a NN which is partially convolutional and partially fully connected or recurrent. In the example below (see table 1 for example), the channel encoder may have a convolutional NN 30a (e.g. five ID convolutional layers followed by a dense layer and a final normalization layer, but a different number of ID convolutional layers could be taken into account). The same also applies to the decoder NN 40a.

[0057] Fig. 3B shows an example of receiver 40' which includes a source-matched channel decoder NN 40'b and a source-matched decoder NN 40'a. This decoder NN 40'b may provide, e.g. through an inverse PEP 46', a representation 24a' of the source signal la. Through a core network 27, the signal 24a" is submitted to another PEP 26', which provides a representation 25', to a channel encoder 30'b and a source-matched channel encoder 30'a. FHIIS24EM36-2024373152. DOCX 12

[0058] Even in this case, the channel encoder NN and channel decoder NN 30'a and 30'b may be formed in a transmitter 20'. The wireless transmission 35 (351), 36 (361) may follow the same principle as above towards a receiver. Each of the receiver 40' and the transmitter 20' may be a RAN (Radio Access Network) station 300, for example.

[0059] The receiver may have a channel decoder 40b and a source-matched channel decoder 40a. A representation of 36 may be provided to an inverse PEP 46. The inverse PEP 46 may provide the representation 46a. The TCP / IP block 44 may provide a further data l'b (5b) which may be, for example, the TCP / IP header, and represent the further data lb. Further, the signal 43a, may be derived from the further data l'b, may be provided to a storage 43. The output 42a of the storage 43 may be provided to a neural network source-matched decryption unit 42. The output of the source-matched decryption 42 may be provided to the source decoder 50. The output of the source decoder 50 may be a signal 5 (e.g. media signal such as video, audio speech, text signal) representing, possibly with a high fidelity, the input data 1, and in particular the source signal la. As explained above, therefore, we can have some decoder apparatus 30 (e.g. at the transmitter) and 20' (e.g., at the network side transmitter) and decoder (e.g. 40 at the user side, and 40' at the network side).

[0060] As explained above, the difference between the channel encoder 30b and the channel encoder implementing the source-matched channel encoder NN 30a is in that the channel encoder 30b is in general not based on any neural network (source-matched or not source- matched), while the source-matched channel encoder 30a, 30'a includes the source- matched neutral network. The same applies in the difference between the channel decoders 40b and 40'b and the source-matched channel decoders 40'a and 40a.

[0061] In the examples of Fig. 3B, the channel decoder apparatus 40 may have the choice at the channel decoder 40b and channel decoder NN 40a by choosing between the stream 140a and the symbols 35 (351), 36 (361). The choice may be based on signally, for example.

[0062] Fig. 3A shows another example of the channel encoder apparatus 30 and the channel encoder apparatus 40. Here, at element 21', a choice is made between a high bitrate path 21a' and a low bitrate path 21b'. FHIIS24EM36-2024373152. DOCX 13

[0063] The high bitrate path 21a' is shown to include the source-matched encryption neural network 22, the storage 23, the TCP / IP block 24, the PEP block 26, as well as the channel encoder NN 30a, while the low bitrate path 21a" has a low bitrate path with a lossy, stronger compression 120, without using the channel encoder NN 30a at all. The stronger compression 120 is a lossy compression which provides a lossly compressed signal 120a. The lossly compressed signal 120a may be provided in the non-source-matched channel decoder 30b. In this case there is not shown the processing of the further data lb which should also be provided to the channel encoder 30b.

[0064] Low bitrate path 21a" may also include a TCP / IP block 44 (not shown) which inserts, for example the TCP / IP header to the lossly compressed signal 120a. Then, the signal 120a is provided to the non-source-matched channel decoder 30b and transmitted to the decoder. The core network side (and in particular the decoder NN 40' and the encoder NN 20') is in Fig. 3A like in Fig. 3C. Instead, the channel decoder apparatus 40, there is the possibility that, based on signaling from the side information to the wireless transmission 35 (351), 36 (361), the channel decoder apparatus 40 determines that a stronger compression 140 is needed from the signal version 140a (which is meant at being the same of the lossly compressed version 120a of the representation 20a of the source signal la). Between the channel decoder NN 40b and the compressor 140 there may be a TCP / IP header reader block (not shown). As can be understood, this can be the low bitrate path 41a". The high bitrate path 41a' of the channel decoder apparatus 40 may include the blocks 46, 44, 43, and / or 42. The output 140a of the compressor 140 may be provided to a storage 41 and inputted, as the representation 40a, to the source decoder. The source decoder 50 will provide the signal 5.

[0065] In examples, when the low bitrate paths 21a" and 41a" are used, the high bitrate paths 21a' and 41a' may be deactivated, and / or when the high bitrate paths 21a' and 41a' are activated, the low bitrate paths 21a" and 41a" may be deactivated.

[0066] The choice 21' may be such that, in case of a transmission which requires high bitrate, the high bitrate path 21a' may be activated, while in the transmissions that require less bitrate, than the low bitrate path 21a" can be activated. The storage 21 may store a multitude of bits for storing the vectorial feature representation 20a: for example, there may be 80 bits stored in the storage 21. In the case of the choice 21' of having chosen the high bitrate FHIIS24EM36-2024373152. DOCX 14 path 21a', then all the bits (or at least a great part of the 80 bits) will be used, as representation 211a to have completely or at least the majority of bits stored in the storage 21. In the case in which the low bitrate path is chosen, then it is possible to select less bits from the storage 21 (e.g. 16). Other than that, a compression 21 can be performed.

[0067] Notably, the compressor 120 has the role of reducing the bitrate, while the compressor 140 has also the role of reducing the bitrate. In case it is not wanted to further reduce the bitrate at the channel decoder apparatus 30, a bypass 140b may be applied.

[0068] Notably, the choice 21a may be upstream than the SM encryption 22, and the provision of the compressed version 14a from the stronger SM compression block 140 may be provided with the lower number of bits, while the version 41a and 21a may be provided in a reduced number of bits. In practice, the signal 120a, 140a is unencrypted.

[0069] Fig. 3C shows another example in which the stronger SM compression 120 is provided, as the compressed signal 120a, (possibly with a TCP / IP header inserted therein, despite not being shown) to a core network, e.g., through a wired connection. Dually, the core network 27 may provide the lossy representation 140a of the source signal la to the stronger compression block 140 (which provides a lossy value 14a'), which is a representation of a low bitrate stored value 211a' at the encoder. Even in this case, the choice between the high bitrate path and the low bitrate path can be made on the basis of the destination of the data, in the case that a wireless transmission is necessary, it may be necessary to choose the high bitrate path 21a' (and / or 41a") and / or in the case of the necessity of wired connection, then it is possible to make use of the low bitrate path 21a" and or 41a". It is to be noted that it is not, in all cases, necessary that at the decoder the same path has been chosen as the encoder.

[0070] In this case, channel encoder apparatus (and more in general, the encoder containing also the source encoder 20) may be, for example, a content media provider for providing a media (e.g. and so on). The source signal la' may be separate among a first, high bitrate signal 211a for a user which requires high quality, and a second signal 211a' to be provided at the low bitrate path 21a" (and obtained by reducing the information from the stored data (for either a user that requires less quality (e.g., a user that does not want to pay for a FHIIS24EM36-2024373152. DOCX 15 premium abonnement) and / or for a user which is connected through a wired connection and, not through the wireless connection).

[0071] In the examples of Fig. 3C, the channel decoder apparatus 40 may have the choice at the channel decoder 40b and channel decoder NN 40a by choosing between the stream 140a and the symbols 35 (351), 36 (361). The choice may be based on signally, for example.

[0072] Fig. 5 shows a signalling between a UE (user equipment) 200 (which may be one of the encoder 2 and the decoder 4) and a RAN (e.g., base station) 300 (e.g. 20' or 40'). The RAN 300 may send a UECapabilityEnquiry 302, which is signalling requesting the UE 200 to send information whether it has the has the channel encoder apparatus (30) with the channelencoder NN (30a). The UE 200 may reply with a UECapabilitylnformation 304 informing that it has or has not the channel encoder apparatus (30) with the channel-encoder NN (30a). Similarly, the RAN 300 may send a UECapabilityEnquiry 302, which is signalling requesting the UE 200 to send information whether it has the has the channel decoder apparatus (40) with the channel-decoder NN (40a). The UE 200 may reply with a UECapabilitylnformation 304 informing that it has or has not the channel deccoder apparatus (40) with the channel-decoder NN (40a). In practice, the UECapabilityEnquiry 302 requests whether the UE has the source match encoding and / or decoding capability, and the UECapabilitylnformation 304 is the response (e.g. binary, e.g. "YES" or "NO"). In case the UECapabilitylnformation 304 is a "NO" response, then the communication continues without the source-matched encoding / decoding. Otherwise, in case the UECapabilitylnformation 304 is a "YES" response, then the communication continues with the source-matched encoding / decoding: for example, the RAN 300 starts sending to the UE 200 the modulation symbols 35, and / or the UE 200 starts sending to the RAN 300 the modulation symbols, in turn. The UECapabilitylnformation 304 from the UE can be NAS (Non-Access Stratum) signalling, for example, or on RRC (Radio Resource Control) layer. In some cases, the UECapabilitylnformation 304 from the UE can be in 5GS session management. In some cases, the UECapabilitylnformation 304 from the UE can be on physical layer. In some cases, the UECapabilitylnformation 304 from the UE can be in control information, DCI (Downlink Control Information), for scheduling of PUSCH (Physical Uplink Shared Channel). In some examples, the UE 200 can send the UECapabilitylnformation 304 autonomously, without receiving the UECapabilitylnformation 304 before. FHIIS24EM36-2024373152. DOCX 16

[0073] Before starting to send the UECapabilityEnquiry 302 and / or the U ECapabilityinformation 304, the RAN 300 and / or the UE 200 may perform measurements on the channel. For example, the UE 200 may perform measurements through Reference Signal Received Power, RSRP, through Reference Signal Received Quality, RSRQ. The UE could send the measured values to the RAN, and to receive a scheduling information based also on the measurement, the scheduling information requesting a particular scheduling and the use or the non-use of the source-matched decoder / encoder NN or not. However, it is also possible that the RAN requests a particular source-matched decoder / encoder NN, i.e. that performs the selection of which NN to use (maintaining the adaptation to the source encoder / decoder).

[0074] It has been understood, indeed, that the neural networks may be not only matched to a specific source encoder / decoder, but also to a specific channel condition. Different instantiations of neural networks may therefore be used for different channels.

[0075] In examples above and below, the encoder-side TCP / IP block 24 may encapsulate the signal 22a and the further data lb in one single stream 24a, and the decoder-side TCP / IP block 44 may de-encapsulate the signal 46a between the header (not shown, but which is further data 5b of l'b) and the media signal 43a.

[0076] Aspects

[0077] Some aspect are discussed here below.

[0078] Aspect 1

[0079] There is provided the channel encoder NN 22, 30a matched to a specific source encoder 20 (e.g. NESC @ 3 kbit / s) or a specific type of source signal (e.g. audio, speech, video) 1, and selection of the specific channel encoder NN 22a, 30 in the transmitter based on the employed specific source encoder 20. Channel decoder NN 10a, 42 matched to a specific source decoder 20 or a specific type of data source 1, and selection of the specific channel decoder NN 42, 40a in the receiver based on the employed specific source decoder or specific type of data source. FHIIS24EM36-2024373152. DOCX 17

[0080] It is possible to select the channel encoder NN 22, 30a and / or the channel decoder 40a, 42 based on either the format of the data to be provided by / to the source encoder / decoder, and / or the specific operating mode of the source encoder / decoder, and / or based on the specific type of source signal 1.

[0081] Aspect 2

[0082] There is provided a combination of the source encoder 20 and the lossy data compressor 120. The data compressor 120 reduces the bitrate of the source encoder output 20a (21a, 211a') for cases, where the data shall not be transferred to a matched channel encoder 30a. This includes the case that a storage 21 is used between source encoder 20 and data compressor 120. Moreover, selection of the specific data compressor 120 in the transmitter based on the employed specific source encoder 20.

[0083] Therefore, the channel encoder apparatus 30 may choose (21') between the high-bitrate path (21a7) and a low-bitrate path (21a"): in the high-rate path (21a7), providing, through the at least one channel-encoder NN (30a), the sequence of symbols (35) from the vectorial feature representation (20a, 21a) in the vectorial format specific of the source encoder (20) or the other feature representation (22a, 23a, 24a, 25), in the low-bitrate path (21a"), providing the vectorial feature representation (20a, 21a) or the other feature representation to a compressor (120), to obtain a lossy- compressed version (120a) of the vectorial feature representation (20a, 21a) or the other feature representation, so as to encode (30a) the lossy-compressed version (120a) of the vectorial feature representation (20a, 21a) or the other feature representation.

[0084] Aspect 3

[0085] There is provided a combination of source-matched channel decoder NN 30a and data compressor 140, where the data compressor 140 reduces the bitrate of the channel decoder output 40a, 41a for cases, where the data shall not be transferred further to a matched channel encoder 30. This includes the case that a storage is used between channel decoder 40 and data compressor 140. Moreover, selection of the specific data compressor 140 in the receiver based on the employed specific source decoder 50. FHIIS24EM36-2024373152. DOCX 18

[0086] Therefore, at the decoder 30 there are the high-bitrate path (41a') and the low-bitrate path (41a"), so that: in the high-bitrate path (41a'), there is provided, through the at least one channeldecoder NN (40b), from the sequence of symbols 36 or the other representation providing information on the sequence of symbols, the vectorial feature representation 20a in the vectorial format specific of the source decoder 20, in the low-bitrate path (41a"), there is provided the vectorial feature representation 14a' or the other feature representation to the lossy compressor 140 not using the at least one channel-decoder NN 40a, to obtain the lossy compressed version of the vectorial feature representation 40a or the other feature representation.

[0087] Aspect 4

[0088] There is provided the „source-matched" encryption device (block) 22 e.g. realized by the neural network that is trained to accept source-encoded original payload data 211a plus random data and to randomize output data, and an associated source-matched decryption device (block) 42 realized by a second neural network that is trained to reproduce the original payload data from the randomized data. It is advantageous that the randomized data has the same ergodic statistical properties as the original payload data but ensures secrecy: an attacker who does not have the source-matched decryption device is not able to reproduce the original payload data from the randomized data.

[0089] Aspect 5

[0090] There is provided a combination of source encoder 20, source-matched encryption device 22, source-matched channel encoder NN 30a, source-matched channel decoder 40a and source-matched decryption device 42, where storage (21, 23, 42, 41) may be introduced after the source encoder (20), or after the encryption device (22), or after the channel decoder NN 30a and / or the decryption device 42. Moreover, a data compressor 140 may be introduced after the decryption device 42 or the storage 41 behind the decryption device 42.

[0091] The storage 21, 23, 43 or 41 may permit to flush bits in case of the non-necessity (e.g. in case of only lower bitrate necessary), or more in general provides a selectable bit rate. FHIIS24EM36-2024373152. DOCX 19

[0092] It is not necessary to have all the storages 21, 23, 43 and 41, and some of them may be avoided.

[0093] Aspect 6

[0094] The performance enhancing proxy 26 introduced between source encoder 20 and the source-matched channel encoder NN 30b that extracts the payload (output from source encoder or from source-matched encryption device) from the input packet and outputs (in 25) as two separate pieces this payload and the remaining parts of the packet (e.g. TCP / IP headers). The performance enhancing proxy 26 can moreover carry out robust header compression (RoHC). The performance enhancing proxy 26 can moreover act as the endpoint of the communication protocols (e.g. TCP / IP) both in direction towards the source (e.g. terminating a TCP / IP connection) as well as towards the sink (e.g. starting a TCP / IP connection). The performance enhancing proxy 26 can serve multiple profiles, e.g. one corresponding to the use of TCP / IP, and further for other transport + network layer protocols.

[0095] More in general, the channel encoder apparatus 30 may insert (in block 24) further data (lb), or part of it, to the vectorial feature representation (20a, 21a) or the other representation (22a), the further data, or part of it, being header information, which is not provided to the channel-encoder NN (30a).

[0096] Therefore, block 24 inserts a TCP / IP header. Notably, the header may be partially „derived" from the source data 20a, 21a, 23a, 24a, e.g. namely the length of the source data packet (e.g. source signal 1) may be a field inside the header, and a checksum (CRC) over the source data may be a field inside the header. Moreover, other protocols may be used in 24 instead of TCP / IP, e.g. UDP / IP or others.

[0097] The fact that block 24 inserts the header (or the further information) lb and the PEP 26 separates it again is advantageous, because it permits to adhere to the standard TCI / IP or UDP / IP.

[0098] The counterpart of is a performance enhancing proxy 46 introduced between source- matched channel decoder NN 40a and source decoder 50 that combines the payload (output from source encoder 20 or from source-matched encryption device 30a) and the remaining FHIIS24EM36-2024373152. DOCX 20 parts of the packet (e.g. TCP / IP headers) 5b, l'b into a packet 46a. This device 26 can moreover act as a counterpart in robust header compression (RoHC). The device 26 can moreover act as the end-point of the communication protocols (e.g. TCP / IP) both in direction towards the source (e.g. terminating a TCP / IP connection) as well as towards the sink (e.g. starting a TCP / IP connection). The proxy counterpart can serve multiple profiles, e.g. one corresponding to the use of TCP / IP, and further for other transport + network layer protocols.

[0099] Aspect 7

[0100] There is provided the partially source-matched channel encoder 30 that is optimized to protect both a payload from a specific source encoder 20 and the non-source-encoded payload (e.g. TCP / IP headers or output of robust header compression) against transmission errors, and selection of the specific channel encoder NN 30a in the receiver based on the employed specific source encoder 20 or specific type of data source.

[0101] There is provided the partially source-matched channel decoder 40 that is optimized to protect both the payload from the specific source encoder 50 and the non-source-encoded payload (e.g. TCP / IP headers or output of robust header compression) against transmission errors, and selection of the specific channel decoder NN 40a in the receiver based on the employed specific source encoder 20 or specific type of data source.

[0102] Aspect 8

[0103] There is provided the combination of the performance enhancing proxy 26 with the partially source-matched channel encoder NN 30a, which receives both the payload 22a or 23a and the remaining parts of the packet (e.g. TCP / IP headers lb) as two separate data pieces. Moreover, the combination of this performance enhancing proxy 26 with the source- matched channel encoder NN 30a and a second channel encoder 30b, where the source- matched channel encoder receives the payload 22a or 23a and the second channel encoder (e.g. without NN) 30b receives the remaining parts of the packet.

[0104] There is provided the combination of the partially source-matched channel decoder NN 40a with the counterpart 46of the performance enhancing proxy 26, which receives both the payload (which will be indicated with 43a or 42a, and which intends to represent the payload 22a or 23a) and the remaining parts of the packet (e.g. TCP / IP headers) (l'b or 5b) as two FHIIS24EM36-2024373152. DOCX 21 separate data pieces. Moreover, a combination of the source-matched channel decoder NN 40a and the second channel decoder 40b with the counterpart 46 of this performance enhancing proxy 26, where the source-matched channel decoder NN 40a outputs the payload 43a and the second channel decoder 40b outputs the remaining parts 5b (l'b)of the packet.

[0105] Aspect 9

[0106] The source-matched channel decoder 40a may generate (or at least use) reliability information (reliability value) for the non-binary features (or more in general for the other information 5b, l'b) in a source-encoded payload. This reliability information conveys to the source decoder 50 the reliability (e.g. a posteriori probability, log-likelihood ratio, LLR) that the associated non-binary feature actually has a specific given / constant value or the value output by the decoder.

[0107] The reliability information may therefore be used for providing different neural networks at the source decoder 50. Or, the source decoder 50 could have the weights parametrized on the reliability values. For example, a high reliability could cause the weights (e.g. in module) to be comparatively high, and a low reliability could cause the weights (e.g. in module) to be comparatively low. The source matched channel decoder may provide reliability information, for example.

[0108] The channel-decoder NN 40a may determine, for at least one symbol or sequence of symbols or at least one value of the other representation, a reliability value associated to the at least one symbol or sequence of symbols or at least one value, so as to provide the reliability value to the source decoder 50. In particular, the channel encoder apparatus 40 (or the device in which the channel encoder apparatus 40 in incorporated) may measure the reliability value from measurements on the occurrence rate of the at least one symbol or sequence of symbols or at least one value of the other representation, for example. In some cases, reliability values (e.g., provided as variances) may be provided, for each component or symbol 45 (or for at least one of them) of the vectorial feature representation 46a, 43a, 42a ,41a, or 40a, as another output value to be provided to the source encoder. The reliability information may be a variance associated with each symbol, for example, or component of the vectorial feature representation 46a, 43a, 42a ,41a, or 40a. FHIIS24EM36-2024373152. DOCX 22

[0109] In alternative or additionally, it may be the channel encoder apparatus 30a which determines, for each symbol or for at least one symbol, the reliability information and may signal it (e.g. it may encode it in a data field in side information of the symbols 35).

[0110] Notably, the reliability value may be provided in a specific field of side information (e.g., encoded by the unmatched channel encoder 30b). Or, the decoder may measure the reliability value from measurements on the occurrence rate of the at least one symbol or sequence of symbols or at least one value of the other representation. The measurements may be processed statistically, to obtain the reliability in terms, e.g. of variance, for example.

[0111] Aspect 10

[0112] Source and (partially) source-matched channel codecs, where the channel decoder apparatus 40 does not only deliver the representation 45 of the symbols 36 in an amount of information which is greater than the amount of information in the channel encoder apparatus 30. For example there may be but real-valued numbers representing parameters from the continuous latent space (infinite alphabet or potentially infinite alphabet) of the symbols 35 as outputted by the source encoder apparatus 40. Once rewritten in binary form (or another quantized form) the amount of information (e.g. number of bits) may be larger than that of the representation 20a, 22a, 23a, 24a, or 25 at the encoder apparatus 30: additional bits are output by the SM channel decoder NN 45 that provide intermediate quantization levels that were not present in the representation 20a, 22a, 23a, 24a, or 25.

[0113] At first, the channel encoder apparatus 30 may generate (and the channel decoder apparatus 40 may decode) the symbols 35 according to a variable alphabet, e.g. with the number of symbols being obtained by training. Different symbols (e.g., at different frequencies) may therefore established, but their number is not necessarily fixed.

[0114] In some cases, it is also possible to generate a continuous latent alphabet, according to which the number of symbols 35, 36 (361) is infinite or at least potentially infinite. The symbols 35, 36 (361) may be defined so as at least one coordinate (e.g. frequency) represents the latent in a continuous, e.g. according to a continuous function which maps the latent value (e.g. 24a, 25) to be encoded by the SM channel encoder NN 30a and the symbol 35 that represents it. The SM channel decoder NN 40a decodes the symbols 36 FHIIS24EM36-2024373152. DOCX 23

[0115] (361) according to a learned continuous function which maps the symbols onto latent values. The function may be obtained by training.

[0116] Further, it may be that the binary informational resolution (e.g. number of bits for quantity of information, and in particular number of bits for value of the feature representation 45, 46a, 43a, 42a, 40a) of the feature representation 45, 46a, 43a, 42a, 40a as generated by the channel decoder apparatus 30 may be variable. In particular, the binary informational resolution of the feature representation 45, 46a, 43a, 42a, 40a may be increased with respect to the binary informational resolution of the representation in the channel encoder apparatus 30. In this way, it is possible to achieve intermediate values from the binarized values.

[0117] Aspect 11

[0118] A chain of (partially) source-matched channel codecs fed by the source encoder apparatus 30, of which at least one channel encoder's input is N bits per packet, and where at least one channel decoder's output is M>N bits per packet, which are then forwarded to the input of the next channel encoder or to the source decoder.

[0119] Aspect 12

[0120] One of the aforementioned schemes, where - additional to the estimated (quantized) nonbinary latents - reliability information about these (quantized) non-binary latents is forwarded from a channel decoder output to either a source decoder input or to the input of another channel encoder. This reliability information conveys to the source decoder or channel encoder the reliability (e.g. a posteriori probability, LLR) that the associated nonbinary feature (i.e. quantized latent) output by the channel decoder actually has a specific value.

[0121] Aspect 13

[0122] (Partially) source-matched channel encoder that accepts - additional to the (quantized) non-binary latents - reliability information at its input.

[0123] Aspect 14

[0124] Joint source and concatenated channel coding: compute joint source and (outer) channel coding JS(O)CC entirely on application layer (i.e. new codec) => code bits from outer FHIIS24EM36-2024373152. DOCX 24 channel encoder would be treated like data and transmitted from CN via RAN to UE; RAN agnostic to JS(O)CC data, would apply its legacy FEC (= inner channel code) encoding on top UE would decode JS(O)CC on application layer.

[0125] Aspect 15

[0126] Aspect based on required control plane signaling for Split JSCC / source-matched channel encoding: Source Encoding on application layer, Channel Encoding in RAN node (gNB) i) NAS signaling (i.e. between UE and CN), e.g. done at Initial Access of UE to network ii) Signaling between BS (gNB) and UE

[0127] (1) UE signals capability to BS: currently for 5G UE capability is specified in 38.331 (RRC Specification)

[0128] (2) network (i.e. RAN) enquires capability from UE, UE sends respective information to the network

[0129] Aspect 16

[0130] Signaling to match source and channel encoder: inform the Physical Layer about what source encoder is used to be able to match the channel encoder (and channel decoder and performance enhancing proxy and its counterpart), see Fig. 10A b) In current 3GPP NR system, the information about the modulation and coding scheme is revealed in Downlink Control Information (DCI) for downlink (BS => UE): in DCI for scheduling of PDSCH (DCI format l_x) c) BS signals to UE used modulation / coding scheme in DCI (currently e.g. format l_0 d) for uplink (UE => BS): in DCI for scheduling of PUSCH (DCI format 0_x): UE reports channel state via CSI reports to BS first, BS is then telling UE in scheduling grant via DCI which modulation / coding scheme to use for uplink e) Inside a UE the information of the used source encoder could possibly be passed from higher layers (application layer) down to the PHY layer.

[0131] It is to be noted above that the encryption NN 22 and the decryption NN 42 may be optional, and in some cases they can be avoided or substituted by non-NN-based encryption / decryption blocks. However, inventors believe that NNs are preferably to be used. FHIIS24EM36-2024373152. DOCX 25

[0132] 1 Discussion

[0133] 1.1 Discussion's Scope

[0134] This document describes the invention "Source-matched channel coding and related communication stack signaling", following the "how to" given in the Fraunhofer patent notification form ("Erfindungsmeldung").

[0135] 1.2 References FHIIS24EM36-2024373152. DOCX 26 FHIIS24EM36-2024373152. DOCX 27

[0136] 2 Detailed Description

[0137] 2.1 Technical Problem Statement (1)

[0138] Multimedia transmission, e.g. speech, comes with various functional blocks in the modern communication systems, including source and channel coding. Conventionally, source coding and channel coding are separated (because it's proven to be therotically optimum for infinite packet length) and their interface is a sequence of bits. Our research results show that if latent values, i.e. numbers, are used instead of bits as the interface and the channel coding scheme is optimized for a specific source codec, e.g. a neural speech codec (see Section 2.6.1), the system achieves significant gains against the conventional approach in terms of perceptual audio quality.

[0139] When a channel coding scheme is tailored to a specific source coding with a specific interface, the capability running such a scheme shall be communicated across the network layers.

[0140] When the ideal interface between such source-aware channel coding scheme has a high bitrate, a data compressor, e.g. a quantizer, the existence and capabilities of such data compressor shall be communicated.

[0141] When a source-aware channel coding needs the meaning / context of the bits transferred through networks layers, e.g. when the channel encoder input is a multidimensional vector whose values are reconstructed using every given number of bits in the data payload, the conventional end-to-end encryption does not work, as the physical layer needs to decrypt the data payload. FHIIS24EM36-2024373152. DOCX 28

[0142] 2.2 Prior Art, Related Patents (2)

[0143] 2.2.1 Conventional Multimedia Transmission

[0144] Conventionally (and in the 3GPP standards), multimedia transmission is implemented (Fig. 6) in the following way: a source encoder 620 generates information bits 620a out of source data 601. A channel encoder unit 630, e.g. encoder of LDPC coding, Polar coding or Turbo coding, (assuming that information bits are independently and identically distributed), generates transmission bits 630' and then a modulation scheme 670 maps the transmission bits 630' to transmission symbols 635. The received symbols 636 (635) are used as an input to the channel decoder, e.g. decoder of LDPC coding, Polar coding or Turbo coding, which computes the estimate of information bits. Finally, the source decoder (through a demodulation block 670, a channel decoder 640 and a source decoder 650) generates the estimate of the source using the estimated information bits. Please see Fig. 6 below for an illustration.

[0145] 2.2.2 Prior Art multimedia transmission in the academic literature

[0146] The advent of deep learning techniques enabled the joint optimization of source coding and speech coding according to semantic / perceptual metrics. This approach achieved significant gains and robustness over the conventional approaches in various multimedia transmission use cases such as text [tl], video [vl], image [il] or speech [si, s2] transmission. Please see Fig. 7 below for a system model of a joint source and channel coding semantic communication. In this case, the input data 601 is directly written, through a semantic extractor 720, into modulation symbols 735, which are then read as modulation symbols 736, which are ingested to a generative mode 740, which in turn provides an output data 705.

[0147] However, to the best of our knowledge, no previous work proposed a separation between source and channel coding when the system is optimized to reduce the semantic / perceptual errors and addressed the potential implementation in the future standards.

[0148] 2.2.3 Neural Speech Codec (NSC)

[0149] NSC [nsc] is a neural network-based speech codec. It comprises an encoder and a respective decoder. The encoder consists of an encoding neural network 820 that compresses a 20 ms speech frame to a 20 dimensional continuous-valued latent vector and a scalar quantizer 830 that quantizes the latent values to quantization levels that can be represented by bits. The decoder side consists of (or more in general comprises) a FHIIS24EM36-2024373152. DOCX 29 dequantizer 840 and a decoding neural network 850. The dequantizer takes the bits as input and reconstructs the latent vector. The decoding neural network takes reconstructed latent vector and syntesise the speech frame accordingly. The entire speech codec is optimized end-to-end to minimize the perceptual audio quality. Please see Fig. 7for an illustration and see the referenced paper for further details.

[0150] 2.2.4 Prior Art Patents

[0151] W023174065 Al - COMMUNICATION METHOD AND RELATED DEVICE

[0152] • signalling (in a PPDU, i.e. possibly targeted towards Wi-Fi) of the use and the parameters of a JSCC method on the Physical Layer from Tx to Rx

[0153] • also signalling information for JSCC from the "source layer"

[0154] US2012155398 AA - SIGNALING TECHNIQUES FOR A MULTIMEDIA-AWARE RADIO AND NETWORK ADAPTATION

[0155] • Make Application Layer parameters accessible to all lower layers (e.g. Physical Layer) in order to improve Quality-of-Experience

[0156] • named Application Layer parameters are: rate distortion function parameters, QoE metrics, multimedia codec type, source encoding and decoding capability, multimedia quality metrics, multimedia coding and layering parameters, frame type, frame rate, quantization parameters, quantity of group of pictures, and applicationlevel constraints, e.g. required latency

[0157] • used, e.g., to prioritize the traffic

[0158] • embodiments: JSCC, QoE-aware link adaptation, resource allocation - all based on cross-layer optimzations

[0159] • source and channel coding is described as separate in the claims

[0160] • the channel encoded signal is generated using one or more application oriented QoE values included in the content specific information

[0161] WO23185389 Al - COMMUNICATION METHOD AND RELATED APPARATUS

[0162] • focus is now on triggering (initiating) such a communication / packet transfer

[0163] 2.3 Solution I Approach (3)

[0164] 2.3.1 Proposed System FHIIS24EM36-2024373152. DOCX 30

[0165] The proposed system in Fig. 1. builds partly on the NSC as explained in section 2.2.3, where an end-to-end trained, neural network-based speech codec is explained. In our proposed system we have used only the neural network elements of the speech encoder and decoder.

[0166] Other works cited in the prior art section optimize the entire system end-to-end to reduce the perceptual loss between source and source estimate. In contrast, in the propsed solution, we optimize the channel encoder and decoder jointly to reduce the distance between the source encoder neural network output and the channel decoder output. In this way, we enable the neural network-based channel coding scheme (e.g. 22, 30a, 40a, 42, 40'a, 30'a) to learn to estimate the transmitted latent vector, that represents the transmitted audio frame, which is the desired input for the source decoder neural network to synthesize the audio frame. When only the channel coding components of the transmission is trained, the interface between source codec and the channel coding remains as the latent vector that is produced by the speech encoder's neural network component.

[0167] 2.3.2 Detailed elaboration of embodiments

[0168] The present technique is also called SplitJSCC.

[0169] Benefits

[0170] • Joint Source Channel Coding (JSCC) can realize gains of up to 10 dB wrt current separated source coding and channel coding schemes in 5G

[0171] • Split JSCC into one part at application layer and one at physical layer allows to transmit inside UE and inside Core Network (CN) - or to store inside a UE or a server - a significantly smaller amount of data compared to the uncompressed source data

[0172] Use Cases

[0173] 1. UE A initiates a source data transfer to UE B

[0174] 2. UE A requests a source data transfer from UE B

[0175] 3. UE A initiates a source data transfer to internet computer B

[0176] 4. UE A requests a source data transfer from internet computer B

[0177] 5. Internet computer B initiates a data transfer to UE A

[0178] 6. Internet computer B requests a source data transfer from UE A

[0179] Potential problems FHIIS24EM36-2024373152. DOCX 31

[0180] 1. added protocol headers (e.g. TCP / IP) with source-independent information that needs to be protected particularly well potential solution: see Data transfer chain o if payload is transparent inside the packets, then no problem; either headers can be protected by specific binary FEC codes, or NN can be trained to protect this binary data very robustly o also consider methods like RoHC to reduce the binary data

[0181] 2. encryption (currently possibly on multiple layers) potential solution: see Data transfer chain

[0182] Mobile Network Control also describe what happens between the layers inside the UE or inside the RAN / CN (see table 4 in Fig. 12)

[0183] Mapping of JSCC functions on Core Network (CN) / Radio Access Network (RAN) *

[0184] • required signaling between NW entities, e.g. from CN to RAN, to signal JSCC, e.g. when JSCC awareness is required in RAN node (gNB) o specific PDU (protocol data unit) / signaling in GTP-U header on user plane between CN / RAN on NG-U (N3) interface o Further stuff possibly involved in CN

[0185] ■ Policy Control Function (PCF)

[0186] ■ QoS Flow specific for JSCC application (i.e. definition of QoS flow parameters like bit rate etc)

[0187] • overview of data rates o Source: 20 symbols I frame = 20x32 (float) = 640 bits I frame

[0188] • JSCC entirely on application layer => modulation symbols out of channel encoder would be treated like data and be transmitted from CN via RAN to UE o RAN agnostic to JSCC data, would apply its legacy FEC encoding on top o UE would decode JSCC on application layer o Signaling in PDU Session setup etc. =>

[0189] ■ Check standard

[0190] • Split JSCC: Source Encoding on application layer, Channel Encoding in RAN node (gNB) o see table below: SplitJSCC (tables 1, 2, 3, in Figs. 10A, 10B, 10C) FHIIS24EM36-2024373152. DOCX 32 o problem: where to cut between source and channel encoding

[0191] ■ transfer of latent space between CN and RAN would involve huge data rate to be transported over fronthaul network

[0192] ■ quantized info bits would have lower data rate o further possiblity: split of Neural Network, but possibly this is even higher bit rate to transfer

[0193] • Transmission of modulation symbols out of channel encoder via 3GPP standard OFDM / SC-FDMA (NR, LTE), TDMA, CDMA, ..

[0194] • PHY layer aspects o channel coding and modulation has to be rate matched to available resources, i.e. Transport Block Size

[0195] Further aspects

[0196] • Life Cycle Management (LCM) of AI / ML models / NNs => 3GPP has defined a framework for this

[0197] Split JSCC*

[0198] Split JSCC: Source Codec on application layer, source-matched Channel Codec in RAN node (gNB)

[0199] • problem: where to cut between source and channel encoding o transfer of latent space between CN and RAN would involve huge data rate to be transported over fronthaul network o quantized info bits would have lower data rate

[0200] • further possiblity: split of Neural Network, but possibly this is even higher bit rate to transfer

[0201] • solution: source encoder produces output either at the minimum possible bitrate or at a higher bitrate than the minimum one that is required by the source decoder; the extra bitrate ("redundancy") just serves to improve the source-matched channel. If not needed, this higher bitrate can be reduced by an extra device ("compressor") to the minimum bitrate required by the source decoder. The higher bitrate needs more storage space (e.g. in platforms like Youtube) than the minimum possible amount, but a certain extra percentage (e.g. 1 / 3 = 33% more) might be acceptable. Similarly, over non-mobile networks (e.g. wireline internet), the higher bitrate will cause more network traffic. FHIIS24EM36-2024373152. DOCX 33

[0202] Focus here: control plane signaling to facilitate Split JSCC

[0203] • NAS signaling (i.e. between UE and CN) o e.g. done at Initial Access of UE to network o capability signaling of UE to support JSCC

[0204] ■ see 24.501 NAS spec

[0205] ■ in 5GMM capability (5GS Mobility Management) Info Element:

[0206] ■ bit / flag indicating 5GMM capability to support (Split)JSCC => 5GMM may be less relevant here

[0207] ■ in 5GSM capability IE (5GS Session Management): indicate UE capability related to PDU sessions

[0208] ■ bit indicating 5GSM capability to support (Split)JSCC

[0209] • Signaling between BS (gNB) and UE o Methods and apparatus: UE capability signaling, e.g. UE is capable of performing JSCC* o UE signals capability to BS: currently specified in 38.331 (RRC Specification)

[0210] ■ network (i.e. RAN) enquires from UE, UE sends respective information to the network

[0211] ■ UECapabilitylnformation is an IE specified in RRC spec

[0212] • Measurement of channel conditions and use o UE can meausre current channel conditions based on reference signals (RS, e.g. SSB, CSI-RS, DM-RS, ...)

[0213] ■ measurement can be done in Downlink (BS => UE) e.g. RSRP (Reference Signal Received Power, in dBm) or RSRQ (Reference Signal Received Quality, in dB)

[0214] ■ measurement can be reported to BS

[0215] ■ BS can use this information to e.g. adapt the channel encoder in the BS for Downlink

[0216] ■ assuming channel reciprocity (i.e. DL measurement also resembles conditions for UL)

[0217] ■ BS can indicate to UE in scheduling grant (after scheduling request by UE for UL) the channel encoder to be used by the UE for UL

[0218] ■ UE autonomously selects channel encoder and embeds respective information together with the FHIIS24EM36-2024373152. DOCX 34 encoded data in UL data for BS so that the BS can decode the data.

[0219] • Channel encoder selection o e.g. pre-configured by network (RAN) for UE, can be a set of multiple possible channel encoders from which a specific one is used for the respective transmissions (has to be signaled, e.g. by BS to UE in scheduling grant)

[0220] Control Flow for data transfer (details of data flow are in section of data transfer chain below)

[0221] • general terms o the input data block to a protocol layer is called Service Data Unit (SDU) o the output data block of a protocol layer is called Protocol Data Unit (PDU)

[0222] • prerequisites o UE has performed Initial Access to the network (see above) o UE has provided to BS the UECapabilitylnformation to BS and further to the Core Network example config (RRC IE) with splitJSCC

[0223] UE-NR-Capability-v2000 : : = SEQUENCE { splitJSCC ENUMERATED

[0224] {supported} OPTIONAL, splitJSCCEncoding ENUMERATED

[0225] {supported} OPTIONAL, split) SCCDecoding ENUMERATED

[0226] {supported} OPTIONAL,

[0227] }

[0228] At any point, where TCP / IP is used, the traffic can be rerouted to some other computer not connected to a mobile network (i.e. not a UE). The TCP connection will then still be host-to-host (e.g. server to UE).

[0229] • As in RoHC, using multiple profiles in the novel PEP, also other transport + network layer protocols can be used besides TCP / IP.

[0230] Whenever the data shall not be transmitted any further (over a mobile communication link), then the (unencrypted) compressed source data can be compressed further by removing FHIIS24EM36-2024373152. DOCX 35 any information that are only relevant for the source-matched channel encoder but not for the source decoder. The position of this source-matched data compressor can be e.g. directly after the source encoder or the storage after the source encoder (if the source data is just consumed on the device or possibly transfered only by ethernet) or it can be in the receiver after the source-matched decryption (if the source data is just consumed on the device, stored for later consumption or possibly transfered only by ethernet).

[0231] 4 Output of the channel decoder

[0232] When the channel encoder receives N bits in conventional channel codecs, then the channel decoder also outputs N bits - or the reliability information about these N bits in the form of a posteriori probabilities (that this bit is actually a 1) or LLRs. In some cases, the channel codec might instead output non-binary symbols (e.g. symbols from the Galois Field GF(256) ) and associated reliability information (e.g. the probability of each of the 256 symbols from GF(256) ). In case the input to the channel encoder are quantized latent values from a source encoder, the bits of the binary representation of the quantized latent values are in conventional channel codecs just like ordinary bits. In some cases, these bits (e.g. 3 bits for a latent value quantized to 8 levels) might be amalgamated by the channel codec to non-binary symbols, e.g. 8 latents might be amalgamated to 3 GF(256) symbols: the 3 bits of latents 1 and 2 and moreover 2 of the bits of latent 3 (i.e. 3+3+2 = 8 bits) go to symbol 1, the remaining bit of latent 3 and the 3 bits of latents 4 and 5 and moreover 1 bit of latent 6 (i.e. 1 +3 +3 + 1 = 8 bits) go to symbols 2, and the remaining 2+3+3 bits of latents 6, 7 and 8 go to symbol 3. But it becomes obvious from this example that there is no direct relationship between the binary representation of the latents (3 bits) and the symbols (8 bits); therefore, it is not straightforward to obtain reliability information about the (quantized) latents from reliability information of the output bits or non-binary output symbols.

[0233] The invented channel codecs always treat the (quantized) latents as integral entities, which are not broken into bits or non-binary symbols for the transmission. Therefore, the natural output of the channel decoder is (quantized) latents and associated reliability information (e.g. probabil ilty of each possible quantized value of the latent). Alternatively, the output might be estimates of the transmitted latent with more quantization levels than were present at the channel encoder's input. This could for instance be the expectation value of the latent. FHIIS24EM36-2024373152. DOCX 36

[0234] Example: let's consider a latent quantized to 8 levels at the channel encoder's input. Before the channel decoder's output, we have the following a posteriori probabilities for each quantization level (the representant is the value representing the respective quantization level):

[0235] The expectation value of the latent would in this case be 11 / 21 = 0.52381.... For approximating this value, a binary representation would need more than 3 bits. E.g. with four bits, the value 10.5 / 21 = 0.5 could be approximated

[0236] 5 Signalling for UE capabilities Unsolved problem:

[0237] • how to inform the Phy Layer about what source encoder is used to be able to match the channel encoder (and channel decoder and performance enhancing proxy and its counterpart) to it? o In current 3GPP NR system, the information about the modulation and coding scheme is revealed in Downlink Control Infomration (DCI)

[0238] ■ for downlink (BS => UE): in DCI for scheduling of PDSCH (DCI format l_x)

[0239] ■ BS signals to UE used modulation / coding scheme in DCI (currently e.g. format l_0

[0240] ■ for uplink (UE => BS): in DCI for scheduling of PUSCH (DCI format 0_x) FHIIS24EM36-2024373152. DOCX 37

[0241] ■ UE would report channel state via CSI reports to BS first, BS is then telling UE in scheduling grant via DCI which modulation / coding scheme to use for uplink o Inside a UE the information of the used source encoder could be passed from higher layers (application layer) down to the PHY layer.

[0242] 6 aspects that can be used for patent embodiments - Data transfer chain

[0243] 1. Channel encoder matched to a specific source encoder (e.g. NESC @ 3 kbit / s) or a specific type of data source (e.g. audio, speech, video), and selection of the specific channel encoder in the transmitter based on the employed specific source encoder. +

[0244] Channel decoder matched to a specific source encoder or a specific type of data source, and selection of the specific channel decoder in the receiver based on the employed specific source encoder or specific type of data source.

[0245] 2. Combination of source encoder and data compressor, where the data compressor reduces the bitrate of the source encoder output for cases, where the data shall not be transfered to a matched channel encoder. This includes the case that a storage is used between source encoder and data compressor. Moreover, selection of the specific data compressor in the transmitter based on the employed specific source encoder.

[0246] 3. Combination of source-matched channel decoder and data compressor, where the data compressor reduces the bitrate of the channel decoder output for cases, where the data shall not be transfered further to a matched channel encoder. This includes the case that a storage is used between channel decoder and data compressor. Moreover, selection of the specific data compressor in the receiver based on the employed specific source encoder.

[0247] 4. A „source-matched" encryption device realized by a neural network that is trained to accept source-encoded original payload data plus random data and to randomize output data, and an associated source-matched decryption device realized by a second neural network that is trained to reproduce the original payload data from the randomized data, where the randomized data has the same ergodic statistical properties as the original payload data but ensures secrecy, i.e. an attacker who FHIIS24EM36-2024373152. DOCX 38 does not have the source-matched decryption device is not able to reproduce the original payload data from the randomized data.

[0248] 5. A combination of source encoder, source-matched encryption device, source- matched channel encoder, source-matched channel decoder and source-matched decryption device, where storage may be introduced after the source encoder, the encryption device, the channel decoder and / or the decryption device. Moreover, a data compressor may be introduced after the decryption device or the storage behind the decryption device.

[0249] 6. A performance enhancing proxy introduced between source encoder and source- matched channel encoder that extracts the payload (output from source encoder or from source-matched encryption device) from the input packet and outputs as two separate pieces this payload and the remaining parts of the packet (e.g. TCP / IP headers). The proxy can moreover carry out robust header compression (RoHC). The proxy can moreover act as the end-point of the communication protocols (e.g. TCP / IP) both in direction towards the source (e.g. terminating a TCP / IP connection) as well as towards the sink (e.g. starting a TCP / IP connection). The proxy can serve multiple profiles, e.g. one corresponding to the use of TCP / IP, and further for other transport + network layer protocols.

[0250] +

[0251] The counterpart of the performance enhancing proxy introduced between source- matched channel decoder and source decoder that combines the payload (output from source encoder or from source-matched encryption device) and the remaining parts of the packet (e.g. TCP / IP headers) into a packet. This device can moreover act as a counterpart in robust header compression (RoHC). The device can moreover act as the end-point of the communication protocols (e.g. TCP / IP) both in direction towards the source (e.g. terminating a TCP / IP connection) as well as towards the sink (e.g. starting a TCP / IP connection). The proxy counterpart can serve multiple profiles, e.g. one corresponding to the use of TCP / IP, and further for other transport + network layer protocols.

[0252] 7. A partially source-matched channel encoder that is optimized to protect both a payload from a specific source encoder and a non-source-encoded payload (e.g. TCP / IP headers or output of robust header compression) against transmission errors, and selection of the specific channel encoder in the receiver based on the employed specific source encoder or specific type of data source. FHIIS24EM36-2024373152. DOCX 39

[0253] +

[0254] A partially source-matched channel decoder that is optimized to protect both a payload from a specific source encoder and a non-source-encoded payload (e.g. TCP / IP headers or output of robust header compression) against transmission errors, and selection of the specific channel decoder in the receiver based on the employed specific source encoder or specific type of data source.

[0255] 8. A combination of the performance enhancing proxy with a partially source-matched channel encoder, which receives both the payload and the remaining parts of the packet (e.g. TCP / IP headers) as two separate data pieces. Moreover, a combination of this performance enhancing proxy with a source-matched channel encoder and a second channel encoder, where the source-matched channel encoder receives the payload and the second channel encoder receives the remaining parts of the packet. +

[0256] A combination of a partially source-matched channel decoder with the counterpart of the performance enhancing proxy, which receives both the payload and the remaining parts of the packet (e.g. TCP / IP headers) as two separate data pieces. Moreover, a combination of a source-matched channel decoder and a second channel decoder with the counterpart of this performance enhancing proxy, where the source-matched channel decoder outputs the payload and the second channel decoder outputs the remaining parts of the packet.

[0257] 9. (Partially) source-matched channel decoder producing reliability information for the non-binary features in a source-encoded payload. This reliability information conveys to the source decoder the reliability (e.g. a posteriori probability, LLR) that the associated non-binary feature actually has a specific given / constant value or the value output by the decoder.

[0258] 10. Source and (partially) source-matched channel codecs, where the channel decoder does not only deliver the input bits of the channel encoder, but real-valued numbers representing parameters from the continuous latent space of the source encoder, quantized and possibly represented by bits, where the number of bits is larger than that of the channel encoder input, i.e. additional bits are output that provide intermediate quantization levels that were not present at the input of the channel encoder.

[0259] 11. A chain of (partially) source-matched channel codecs fed by a source encoder, of which at least one channel encoder's input is N bits per packet, and where at least FHIIS24EM36-2024373152. DOCX 40 one channel decoder's output is M>N bits per packet, which are then forwarded to the input of the next channel encoder or to the source decoder.

[0260] 12. One of the aforementioned schemes, where - additional to the estimated (quantized) non-binary latents - reliability information about these (quantized) nonbinary latents is forwarded from a channel decoder output to either a source decoder input or to the input of another channel encoder. This reliability information conveys to the source decoder or channel encoder the reliability (e.g. a posteriori probability, LLR) that the associated non-binary feature (i.e. quantized latent) output by the channel decoder actually has a specific value.

[0261] 13. (Partially) source-matched channel encoder that accepts - additional to the (quantized) non-binary latents - reliability information at its input.

[0262] 14. Joint source and concatenated channel coding: compute joint source and (outer) channel coding JS(O)CC entirely on application layer (i.e. new codec) => code bits from outer channel encoder would be treated like data and transmitted from CN via RAN to UE; RAN agnostic to JS(O)CC data, would apply its legacy FEC (= inner channel code) encoding on top

[0263] UE would decode JS(O)CC on application layer

[0264] 15. Claims based on required control plane signaling for Split JSCC / source-matched channel encoding: Source Encoding on application layer, Channel Encoding in RAN node (gNB) i) NAS signaling (i.e. between UE and CN), e.g. done at Initial Access of UE to network ii) Signaling between BS (gNB) and UE

[0265] (1) UE signals capability to BS: currently for 5G UE capability is specified in 38.331 (RRC Specification)

[0266] (2) network (i.e. RAN) enquires capability from UE, UE sends respective information to the network

[0267] 12. Signaling to match source and channel encoder: inform the Phy Layer about what source encoder is used to be able to match the channel encoder (and channel decoder and performance enhancing proxy and its counterpart) f) In current 3GPP NR system, the information about the modulation and coding scheme is revealed in Downlink Control Information (DCI) for downlink (BS => UE): in DCI for scheduling of PDSCH (DCI format l_x) g) BS signals to UE used modulation / coding scheme in DCI (currently e.g. format l_0 FHIIS24EM36-2024373152. DOCX 41 h) for uplink (UE => BS): in DCI for scheduling of PUSCH (DCI format 0_x): UE reports channel state via CSI reports to BS first, BS is then telling UE in scheduling grant via DCI which modulation / coding scheme to use for uplink i) Inside a UE the information of the used source encoder could possibly be passed from higher layers (application layer) down to the PHY layer.

[0268] 6.1 Assessment of the Invention

[0269] 6.1.1 Benefits of the Invention (4a)

[0270] • Significant gain in terms of achieved perceptual audio quality with the given transmit power (or significant power gain for the target perceptual audio quality).

[0271] • The robustness of the system against varying noise levels is increased, i.e. graceful degradation of the audio quality by reducing signal strength.

[0272] • Spectral efficiency of the transmission scheme is increased.

[0273] • The source-match channel coding approach sustains the benefits of JSCC approaches while enabling e separation between source and channel coding.

[0274] 7. Source-Matched Channel Coding for Semantic Speech 1 1 dismission

[0275] This section proposes an example of novel machine learning-based source-matched channel coding approach for semantic transmission of short speech frames based on examples above. We use prior -art speech codec to compare conventional channel coded transmission, uncoded transmission, and the proposed scheme. Our results demonstrate that our separate source and channel coding scheme for short frames achieves performance gains similar to those achievable in joint source and channel coding (JSCC) methods or even surpasses them. By keeping separated source and channel coding, we take a step towards addressing network-related aspects and reduce the training complexity of the transmission system. Additionally, we present results on peak-to-average power ratio (PAPR) constrained transmission to facilitate the implementation of the proposed approach in real-world applications.

[0276] 7.1 Introduction

[0277] The ultimate goal of communication is the exchange of semantic information, i.e., information relevant to the receiving user, e.g., knowledge or meaning. Conventional communication systems separate source and channel coding functions, where source coding aims at representing the source information in the most compact form possible into FHIIS24EM36-2024373152. DOCX 42

[0278] (binary) symbols I bits, whereas channel coding encodes these symbols - agnostic to the used source coding - for a reliable transmission over the per se unreliable communication channel. For infinite block length source and channel codes and with the objective of perfect (i.e., error-free) reconstruction, this separation is proven to be theoretically optimum. However, in practical communication systems and with short block lengths, the separation is sub-optimum. Therefore, joint source and channel coding (JSCC) for semantic transmission (like in Fig. 7) has attracted considerable research interest recently, and this has been further fueled by the advent of deep learning (DL), facilitating the realization and optimization of communication systems according to semantic objectives. Since mobile data traffic currently represents a major contribution to the global energy consumption and is expected to further grow exponentially, semantic communication is a promising paradigm for more resource-efficient transmission [8].

[0279] Semantic communication systems have been explored across various domains using deep- learning(DL)-based JSCC approaches including video, image, text and speech transmission [5], [1]. Corresponding works have demonstrated significant achievable performance gains of the proposed schemes, typically measured in terms of perceptual / semantic evaluation metrics over signal-to-noise ratio (SNR), when compared to conventional systems in their respective fields. Moreover, JSCC methods that are designed for semantic objectives typically exhibit a graceful degradation in performance with deteriorating channel quality, which highlights the robustness of this approach under variable transmission conditions

[0280] In this work, we focus on the semantic transmission of speech signals by employing a pretrained neural speech codec (NSC) [2] (cf. Section 3.2.2) (e.g. at least one of the NNs 30a, 40a, and / or 22, 42) as the source coding component (e.g. to be implemented in one of the apparatuses 30, 40). We include a separate channel coding scheme that is matched o this source codec. Therefore, we refer to our approach as source-matched channel coding. Hence, our proposed scheme bridges the gap between conventional JSCC pf Fig. 7 with its significant gains and the complete separation of source and channel coding. This allows, e.g., to store source-encoded data (e.g. audiobooks) on servers and transmit them on- demand to users using the source-matched channel codec, or to transmit a source-encoded payload via multiple hops (e.g.. mobile phone to base station, backhaul and remote base station to remote phone) - each individually protected by source-matched channel coding while the source coding remains end-to-end.

[0281] Our results demonstrate that source-matched channel coding achieves performance gains comparable to or even surpassing those observed for JSCC approaches. Furthermore, it consistently outperforms conventional transmission schemes that rely on the same source FHIIS24EM36-2024373152. DOCX 43 codec, emphasizing the advantages of tailoring channel coding to the source codec. Our findings also reveal that the output bitrate of source encoded data can be reduced to levels similar to those in conventional systems. Finally, we demonstrate that the proposed approach can operate under peak-to-average power ratio (PAPR) constraints, highlighting its practical feasibility for real-world applications.

[0282] 7.2 System Model

[0283] In this section, we introduce a system model that can represent both the considered benchmark schemes and the proposed scheme. The system model comprises a transmitter (see Fig. 11), a channel, and a receiver (see Fig. 12). The purpose of the transmission system is to convey speech at a constant frame duration. In practise, Figs. 11 and 12 show models which are valid for both the invention and the prior art.

[0284] 7.2.1 Transmitter (e.g., 30, 20' in the invention)

[0285] A sequence of equi-distant samples of a speech record is organized in vector s e ]R^, where lsdenotes the number of samples per audio frame (s can be the source signal 1 or la in its original version, upstream to the source encoder 20). The source encoder 20 maps the given audio frame vector s (1 or la) to a feature vector f (which may represent the vectorial feature representation 20a, or the other representation 20a, 21a, 22a, 23a, 24a, 25), f = s_enc(s) e IR , (1) where s_enc(-) and lfdenote the encoding operation and the length of feature vector (20a, 21a, 22a, 23a, 24a, 25), respectively, f may comprise binary symbols in conventional schemes, however, we allow for real-valued symbols in order to include our proposed scheme as well. The block responsible for modulation and coding referred to as mod_cod(-) (channel encoder apparatus 30) may accept f as input (e.g. 25) and generate the real- valued encoded transmit symbol vector a (35), a = mod_cod( ) e ]Ra, (2) where ladenotes the number of transmit symbols per audio frame. Subsequently, pulse shaping (not shown in Figs. 3A-3C as being implicit) may be applied to the transmit symbol vector a, yielding the equivalent baseband representation x(t) of the continuous-time transmit signal (351), FHIIS24EM36-2024373152. DOCX 44 where T denotes the symbol period, and g(t) (implemented in block 350) is a pulse shaping filter satisfying the first Nyquist criterion.

[0286] 7.2.2 Channel

[0287] We assume a transmission over an additive white Gaussian noise (AWGN) channel with a receiver with perfect carrier synchronization (35aa in Figs. 1 and 2). Then, the continuoustime received signal r(t) (361) in equivalent baseband can be expressed as r(t) = x(t) + n(t), (4) where n(t) is white Gaussian noise (35' in Figs. 1 and 2) with double-sided power spectral density No.

[0288] 7.2.3 Receiver

[0289] In the receiver (e.g. 40, 40'), after matched filtering and symbol-spaced sampling (in block 360), assuming perfect clock synchronization, a discrete-time received sequence with samples y[ / c] (36) collected in vector y e is obtained from the signal 361, so that

[0290] Thus, the output of the discrete-time channel can be modeled as y = a + n, where n e denotes the noise vector comprising the discrete-time noise components. The block responsible for demodulation and channel decoding characterized by the function dem_dec(- ) may accept y as input and delivers an estimate of the feature vector, f = dem_dec(y) e JR / . (6)

[0291] Furthermore, the source decoder 50 with law s_dec(-) may produce an estimate 5 of the audio frame based on the estimated feature vector, s = s_dec( ) e ]Ris. (7)

[0292] 7.3 Reference Schemes

[0293] In this section, we provide our preferred prior-art channel and speech coding processes that serve as reference components as well as the transmission benchmark schemes in our study for comparison with our proposed approach, as detailed in Section V. FHIIS24EM36-2024373152. DOCX 45

[0294] 7.3.1 Channel Coding

[0295] 7.3.1.1 Turbo Coding

[0296] For performance comparisons, turbo coding with 3GPP interleaver, a code rate of 1 / 3, constraint length of 4, and generator polynomials of 15 and 13 in octal notation are selected for conventional channel coding due to its short packet transmission performance. The Nvidia Sionna library [4] is used for implementation of turbo coding.

[0297] 7.3.2 Speech Coding

[0298] 73.2.1 Enhanced Voice Services (EVS)

[0299] We consider the 3GPP Enhanced Voice Services (EVS) encoder [3] as a baseline legacy speech codec. In our study, we operate EVS with 7.2 kbps and packet concealment.

[0300] 73.2.2 Neural Speech Codec (NSC)

[0301] The recently proposed real-time capable NSC [2] achieves low processing latency and computational complexity, making it suitable for real-time communication applications. It employs a neural network (NN), SENN, which encodes 20 ms audio frames sampled at 16 kHz (Zs= 320) into a latent space lre ]R20. The latents are discretized using 3-bit scalar quantization (SQ) per dimension, producing a bitstream b e {O,l}60. On the decoder side, a corresponding dequantization step is applied, followed by reconstruction through a decoding NN, SDNN. For further details, refer to [2].

[0302] 7.3.3 Transmission Benchmark Schemes

[0303] For the comparison with our proposed scheme, we have chosen the following combinations of source and channel codecs:

[0304] 7.3.3.1 EVS + Turbo

[0305] Here, we apply EVS with 7.2 kbps bit rate for source coding and turbo coding with specifications according to Section 3.1.1 for channel coding. Afterwards, the channel- encoded bits are mapped to binary phase-shift keying (BPSK) symbols. At the receiver side, the decoding counterparts of the encoders are employed.

[0306] 7.3.3.2 NSC + Turbo

[0307] This scheme is similar to EVS + Turbo, except that we adopt NSC with 3.0 kbps for the source coding. FHIIS24EM36-2024373152. DOCX 46

[0308] 73.3.3 NSC-L Uncoded

[0309] Here we use the SENNmodule of NSC for source coding. This transmission scheme does not use an explicit channel codec - instead, the latent representation extracted by the source encoder is first normalized by subtracting its mean and scaling it for unit variance (with mean and variance computed based on the training set of VCTK [7]) and then forwarded (as a real-valued sample) to the transmit pulse shaping unit. After reception, the dem_dec module scales the received signal to attain the latent representation's original variance and adds its mean. Speech is reconstructed using the resulting latent representation values as input to the SDNNmodule of NSC. With this scheme, we aim at exploring the graceful degradation property of the source codec.

[0310] 7.3.3.4 Deep Joint Source-Channel Analog Coding for Low-Latency Speech Transmission [1]

[0311] This scheme is included to evaluate whether the performance of the proposed approach is comparable to that of other schemes optimized specifically for perceptual audio quality. The system model and evaluation metrics are compatible to the framework used in this study, and thus a fair comparison can be guaranteed.

[0312] 7.4 Proposed Scheme with Source-Matched Channel Coding

[0313] In this section, we introduce the system model components of the proposed scheme and the corresponding training procedure. NNs are utilized for each functional block.

[0314] 7.4.1 System Model Components

[0315] The NN block of NSC which maps the speech frame to a latent space representation, SENN, and the block generating the speech frame from the reconstructed latent vector, SDNN(see Fig. 8 for an illustration), are used as source encoder s_enc and source decoder s_dec, respectively. The model we used has ls= 320, = 20, Zb= 60. The source encoder and decoder networks (e.g. 22, 42, 30a, 40a) may be pre-trained only for the purpose of the present comparison. Details regarding their architecture and training procedure are available in [2]. An NN, CENN, is employed in the mod_cod unit and another NN, CDNN, is utilized in the dem_dec unit.

[0316] In contrast to conventional channel codecs, the NN (30a) responsible for channel encoding, CENN, accepts real-valued latents as its input rather than binary symbols / bits. Its architecture is inspired from [6], and summarized in Table 1. It may comprise five ID convolutional layers followed by a dense layer and a final normalization layer. Each convolutional layer may use a kernel size of five, with zero-padding on both sides to FHIIS24EM36-2024373152. DOCX 47 maintain the input dimension. The stride, dilation and groups are chosen to one. This means that each filter processes every input channel individually without skipping or dilating values, and each input channel is connected to every output channel. The first convolutional layer operates with one channel, while the subsequent four layers may have e.g. 50 input channels each (or a number between 30 and 70, in some examples). In all layers except the normalization layer a bias is applied, and the exponential linear unit (ELU) activation function is adopted. The dense layer reduces the output to a single channel, enabling the network to produce the transmit vector that comprises the transmit symbols. We have considered two power normalization modes, both of which make sure that the long term average symbol energy is one. In the first mode, the system maintains a constant average transmit packet energy while allowing for dynamic energy allocation between packets. This enables the encoder to assign higher energy to packets corresponding to perceptually important speech frames and lower energy to less significant ones. We name this mode as Dynamic Packet Energy Mode (DPEM). In the second mode, a constant packet energy constraint is enforced regardless of the content of the packet, that is named as Constant Packet Energy Mode (CPEM).

[0317] Additionally, we have evaluated a scheme with a delimiter operation at the 8thlayer (which provides the transmit symbols) to account for a PAPR constraint. FHIIS24EM36-2024373152. DOCX 48

[0318] The NN 40a responsible for channel decoding, CDNN, may comprise ID convolutional layers of the same type as CENN(30a). The final layer is a dense layer with no activation. Its architecture is summarized in the lower part of Table 1.

[0319] 7.4.2 Training

[0320] CENN(30a) and CDNN(40a) are jointly trained end-to-end. The training chain may begin with the latent representation 20a delivered by SENN(20) as the input to CENN(or with another vectorial feature representation such as 21a, 22a, 23a, 24a, 25) and ends with the latent representation estimate 45 (or another vectorial feature representation such as 46a, 43a, 42a, 41a, 40a) at the output of CDNN(40a). The training dataset comprises latent representations generated by SENN(20) from speech recordings in the VCTK training set. For each recording, latent representations are extracted for each audio frame of duration TF= 20 ms, which are then stacked and shuffled to create the final training dataset, that is denoted by T>. The channel encoder NN 30a and channel decoder NN 40a may be trained end-to-end to minimize a composite loss function defined as where £i(-,-) and £2( ) denote the mean absolute error (MAE) and mean squared error (MSE), respectively. We empirically select A = 0.5. We linearly combine MSE and MAE because using only MSE heavily penalizes large errors and tolerates smaller deviations around the true level, while adding MAE results in a more balanced error distribution. The training process uses Ep / N0levels uniformly sampled (in dB domain) from the interval 12dB < Ep / N0< 20dB, where Epdenotes the average transmit energy per audio frame. The NN weights are updated using the Adam optimizer with a learning rate of 0.001 and batch size of 3200, p values of 0.9 and 0.999, and no weight decay.

[0321] For the DPEM, batch normalization is applied during training to generate transmit symbols. After convergence, the weights and the normalization parameters, calculated over the entire speech records in the VCTK training set, are frozen. The fixed average mean and variance values (that became a part of the model) are used for inference. In contrast, in the CPEM, every packet is normalized individually both in training and inference.

[0322] 7.5 Numerical Results and Discussion

[0323] This section presents our experiments, the used metrics, and a performance comparison. FHIIS24EM36-2024373152. DOCX 49

[0324] 7.5.1 Metrics

[0325] All records in the VCTK testing set are used to extract audio frames to be transmitted. All schemes considered for the numerical results aim to convey a 20 ms speech frame sampled at 16 kHz. While we transmit one packet per audio frame, the different schemes yield different transmit packet lengths ladue to the differences in the compression rate of the source coding and the modulation and coding rate of the channel coding scheme (cf. Table 2). Also the transmit symbol rates Rsym(symbols over the transmission channel) differ.

[0326] Table 2: Transmit Packet Lengths and Symbol Rates for TF= 20 ms

[0327] We observe that the proposed scheme uses the least number of transmit symbols and achieves thus the highest bandwidth efficiency. By increasing the number of transmit symbols, we can expect a further improvement in power efficiency, i.e., a lower required Ep / N0for a certain target performance Similarly, if we replace BPSK by higher-order modulation for EVS+Turbo and NSC+Turbo, the required Ep / N0will grow further.

[0328] For performance assessment, we employ the Extended Short-Time Objective Intelligibility (ESTOI) Score which is a well-known and broadly adopted objective speech quality measure with high correlation to the intelligibility of speech signals of varying quality.

[0329] 7.5.2 Discussion

[0330] Fig. 13 presents the perceptual quality (without a physical layer transmission) of the EVS and NSC speech codecs at specified bit rates and shows the ESTOI score versus Ep / N0over the AWGN channel. An ESTOI score of approx. 0.7 corresponds to an acceptable quality and serves as our reference threshold. Comparing EVS + Turbo and NSC + Turbo, a FHIIS24EM36-2024373152. DOCX 50 reduction in source coding rate from 7.2 to 3 kbps achieves approximately a 3 dB improvement at the reference threshold. Surprisingly, NSC-L Uncoded outperforms NSC+Turbo when the ESTOI score is 0.76 or lower. This improvement likely arises from two main factors: first, NSC employs bounded uniform noise during training to approximate quantization effects, which not only enhances quantization performance but also introduces inherent forward error correction capabilities in SENN. Furthermore, uncoded transmission over AWGN benefits from graceful degradation.

[0331] At the reference threshold, our proposed scheme shows a gain of approximately 4.5 dB over NSC-L Uncoded, approx. 7.5 dB over NSC+Turbo, and roughly 11 dB over EVS+Turbo, demonstrating the benefits of a source-matched channel coding approach, as the source codec is identical to that of NSC+Turbo (except for the discarded quantization and dequantization modules). Conventional Shannon-based channel coding focuses on error- free transmission for SNRs above a target SNR, which manifests itself in the well-known "turbo cliff - here on the ESTOI performance EVS+Turbo and NSC+Turbo. By contrast, our proposed source-matched channel codec attempts to provide best reconstruction of the latent representation for the complete SNR range used in the training, resulting in graceful degradation.

[0332] The widely-used Perceptual Evaluation of Speech Quality (PESQ) score does unfortunately not provide meaningful results for generative Al-based codecs like NSC. However, ESTOI scores are available for the JSCC scheme proposed in [1]. As the speech frame duration in [1] is only TF= 8 ms, while ours is 20 ms, we introduce the useful energy Elms / N0received during the transmission of one millisecond c speech for a fair comparison: for our scheme • 1 ms / / V0, whereas for the scheme in [1], we have Elms / N0= here the used SNR for a real-valued transmission corresponds to Es / (W0 / 2), n is the audio frame length (in samples) and a = la / n with transmit packet length la. Fig. 14 shows a comparison of the proposed scheme and the scheme of [1] (curves for the parameter sets from the paper TF= 8 ms, n = 128, SNR = 0 dB and 10 dB, p = 100). Moreover, it shows the required symbol rates Rsymof the schemes.

[0333] It becomes apparent that for acceptable ESTOI scores above 0.7, our scheme gains up to 10 dB upon the JSCC scheme of [1], even though our scheme offers the additional advantage of separate source and channel coding and the scheme of [1] uses globally optimized source and channel coding parameters. Moreover, the bandwidth efficiency of our proposed scheme is higher. FHIIS24EM36-2024373152. DOCX 51

[0334] In the upper part of Fig. 15, the performance of the proposed system is shown under different transmit energy / power constraints. In particular, the impact of DPEM and CPEM is investigated. Additionally, the system is evaluated when constraining the PAPR to 4, 3, and 2 (in linear scale). A PAPR of 3 is particularly relevant as it corresponds to that for M- QAM / M-ASK constellations for M -> oo. Here, a slight performance degradation of approximately 0.02 in ESTOI score results compared to the model without PAPR constraint.

[0335] Next we consider the effect of handing over 3-bit integer values ("Int3") insmtead of 32-bit floating-point ("Float32") at the interfaces between source and channel codecs. The lower part of Fig. 15 shows rather small performance differences of 0.02 to 0.03 ESTOI, where the left label Int3ov F / oat32) represents the interface between source and channel encoder and the right one that between channel and source decoder. Hence, the NSC's standard output (with 3 bits per latent representation) can be used with only a minor performance penalty.

[0336] The observed gains over comparable schemes might be attributed to the absence of unequal-error-protection (UEP) in the turbo code and residual redundancy in the source- encoded latent vectors. Even though the employed NSC source codec is among the currently most efficient ones and hence supposed to reduce any redundancy as much as possible, we found by inspection of the encoder output that it still contains residual redundancy. This redundancy is reflected in statistical depencencies between the entries of the latent vector, and in a non-uniform probability distribution of the latent representations.

[0337] Summarizing, the proposed technique permits a better graceful degradation in case of lower quality of the transmission.

[0338] 7.6 Conclusion in paricuarl for the speech coding

[0339] We have found that it is possible to separate source and channel coding while realizing similar gains as with joint source-channel coding (JSCC) approaches. Our channel codec is neural-network-based and optimized specifically for an input that is a latent space representation of the speech produced by a prior-art speech encoder - which we refer to as "source-matched channel coding". By focusing on channel coding optimization alone, the neural network training is simplified, as the source coding parameters can remain fixed. This approach also allows source coding to remain on the application layer of the transmission network, such that source-encoded, i.e. compressed, data can be stored on servers instead of raw source data. While the source coding and decoding is only carried out on both ends of the transmission chain, in a multi-hop transmission our channel coding FHIIS24EM36-2024373152. DOCX 52 can be applied at each hop's physical layer. To enable compatibility with traditional communication protocols like TCP / IP and paradigms like end-to-end encryption, further research is needed to look into, e.g., source-matched encryption.

[0340] 7.7 Further characterization of some figures

[0341] Fig. 11: Representation of transmitter.

[0342] Fig. 12: Representation of receiver.

[0343] Fig. 3: NSC encoder and decoder block diagram.

[0344] Fig. 13: Comparison of the benchmark schemes.

[0345] Fig. 14: ESTOI score for the proposed scheme and the best performing scheme of [1] (purple and red curve for varying a). The numbers along the curves represent the transmit symbol rates Rsym(in kHz).

[0346] Fig. 15: Performance with PAPR constraints (upper) and quantized source-channel codec interface (lower).

[0347] References

[0348] [1] Mohammad Bokaei, Jesper Jensen, Simon Doclo, and Jan A~stergaard. Deep joint source-channel analog coding for low-latency speech transmission over gaussian channels. In 202331st European Signal Processing Conference (EUSIPCO), pages 426- 430, 2023.

[0349] [2] Andreas Brendel, Nicola Pia, Kishan Gupta, Lyonel Behringer, Guillaume Fuchs, and Markus Multrus. Neural speech coding for real-time communications using constant bitrate scalar quantization. 2024.

[0350] [3] Martin Dietz, Markus Multrus, Vaclav Eksler, et al. Overview of the evs codec architecture. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5698-5702, 2015.

[0351] [4] Jakob Hoydis, Sebastian Cammerer, FayA§al Ait Aoudia, Avinash Vem, Nikolaus Binder, Guillermo Marcus, and Alexander Keller. Sionna: An open-source library for nextgeneration physical layer research, 2023.

[0352] [5] Nazmul Islam and Seokjoo Shin. Deep learning in physical layer: Review on data driven end-to-end communication systems and their enabling semantic applications. IEEE Open Journal of the Communications Society, 5.

[0353] [6] Yihan Jiang, Hyeji Kim, Himanshu Asnani, Sreeram Kannan, Sewoong Oh, and Pramod Viswanath. Turbo autoencoder: Deep learning based channel codes for point-to- point communication channels, 2019.

[0354] [7] C. Veaux J. Yamagishi and K. MacDonald. Cstr vctk corpus: English multispeaker corpus for cstr voice cloning toolkit, 2019. FHIIS24EM36-2024373152. DOCX 53

[0355] [8] Wen Tong and Geoffrey Ye Li. Nine challenges in artificial intelligence and wireless communications for 6g. IEEE Wireless Communications, 29(4): 140-145, 2022.

[0356] Further examples

[0357] Generally, examples may be implemented as a computer program product with program instructions, the program instructions being operative for performing one of the methods when the computer program product runs on a computer. The program instructions may for example be stored on a machine readable medium. Other examples comprise the computer program for performing one of the methods described herein, stored on a machine readable carrier. In other words, an example of method is, therefore, a computer program having a program instructions for performing one of the methods described herein, when the computer program runs on a computer. A further example of the methods is, therefore, a data carrier medium (or a digital storage medium, or a computer-readable medium) comprising, recorded thereon, the computer program for performing one of the methods described herein. The data carrier medium, the digital storage medium or the recorded medium are tangible and / or non-transitionary, rather than signals which are intangible and transitory. A further example of the method is, therefore, a data stream or a sequence of signals representing the computer program for performing one of the methods described herein. The data stream or the sequence of signals may for example be transferred via a data communication connection, for example via the Internet. A further example comprises a processing means, for example a computer, or a programmable logic device performing one of the methods described herein. A further example comprises a computer having installed thereon the computer program for performing one of the methods described herein. A further example comprises an apparatus or a system transferring (for example, electronically or optically) a computer program for performing one of the methods described herein to a receiver. The receiver may, for example, be a computer, a mobile device, a memory device or the like. The apparatus or system may, for example, comprise a file server for transferring the computer program to the receiver. In some examples, a programmable logic device (for example, a field programmable gate array) may be used to perform some or all of the functionalities of the methods described herein. In some examples, a field programmable gate array may cooperate with a microprocessor in order to perform one of the methods described herein. Generally, the methods may be performed by any appropriate hardware apparatus. The above described examples are merely illustrative for the principles discussed above. It is understood that modifications and variations of the arrangements and the details described herein will be apparent. It is the intent, therefore, to be limited by the scope of the claims and not by the specific details FHIIS24EM36-2024373152. DOCX 54 presented by way of description and explanation of the examples herein. Equal or equivalent elements or elements with equal or equivalent functionality are denoted in the following description by equal or equivalent reference numerals even if occurring in different figures.

[0358] In examples above there is provided, inter alia, a channel decoder apparatus (e.g. 40) for generating a vectorial feature representation as a vectorial feature representation (e.g. 40a, 45) in a format specific of a source decoder (e.g. 50), from a sequence of symbols (e.g. 36), the vectorial feature representation (e.g. 40a) representing a source signal (e.g. 1, 5), the channel decoder apparatus being configured to apply at least one channeldecoder neural network, NN (e.g. 40a, 42), to the sequence of symbols (e.g. 36) or another representation (e.g. (e.g. e.g. 46a, 43a, 42a) of the sequence of symbols, thereby generating the vectorial feature representation in the format specific of the source decoder (e.g. (e.g. e.g. 50).

[0359] The at least one NN (e.g. 40a, 42) may be a fully-connected NN or is a recurrent NN which, when developed, is a fully connected NN.

[0360] The at least one NN (e.g. 40a, 42) may be such that the NN is not separable in multiple sub-NNs independent from each other.

[0361] The channel-decoder NN may process the sequence of symbols or the other representation, through a sequence of layers using weights, the weights weighting, in parallel with each other, the sequence of symbols or the other representation.

[0362] The channel decoder apparatus may perform a selection selecting, among a plurality of selectable channel-decoder NNs, a selected at least one channel-decoder NN based at least in part on the specific source decoder.

[0363] The channel decoder apparatus may perform a selection selecting, among a plurality of selectable channel- decoder NNs, a selected at least one channel-decoder NN based at least in part on the vectorial format specific of the source decoder (e.g. 20). FHIIS24EM36-2024373152. DOCX 55

[0364] The channel decoder apparatus may perform a selection selecting, among a plurality of selectable channel-decoder NNs, a selected at least one channel-decoder NN based at least in part on a specific operating mode of the source decoder.

[0365] The channel decoder apparatus may perform a selection selecting, among a plurality of selectable channel- decoder NNs, a selected at least one channel-decoder NN based at least in part on a specific type of source signal.

[0366] The channel decoder apparatus may perform the selection based on an indication from the source decoder (e.g. 20).

[0367] The channel decoder apparatus may perform the selection based on a request from a transmitter.

[0368] The channel decoder apparatus may perform the selection based on a request from a higher layer.

[0369] The channel decoder apparatus may receive, from a transmitter, a request for indicating whether the channel decoder apparatus has the capability of performing the selection, and may respond that the channel decoder apparatus has the capability of performing the selection.

[0370] The channel decoder apparatus may be carried out, together with a transmitter, a handshaking procedure defining, based at least in part on the request from the receiver, on the source decoder, and on selectable NNs, the at least one selected NN.

[0371] The channel decoder apparatus may receive, from a higher layer, a request for indicating whether the channel decoder apparatus has the capability of performing the selection, and to respond that the channel decoder apparatus has the capability of performing the selection.

[0372] The selection may select the at least one selected NN (e.g. 40a) irrespective of the channel conditions. FHIIS24EM36-2024373152. DOCX 56

[0373] The selection may selectat least one selected NN (e.g. 40a) based, at least partially, on the measurements on the channel.

[0374] The channel decoder apparatus may perform at least one of the measurements on the channel.

[0375] The implemented in a user equipment, UE, the channel decoder being configured to perform the at least one of the measurements through Reference Signal Received Power, RSRP.

[0376] The channel decoder apparatus may be implemented in a user equipment, UE, the channel decoder being configured to perform the at least one of the measurements through Reference Signal Received Quality, RSRQ.

[0377] The channel decoder apparatus may be implemented at a user equipment, UE, report, to a radio access network, RAN, side the measurements, subsequently waiting for an indication, in the scheduling, of the selection to be performed.

[0378] The channel decoder apparatus may be implemented at a radio access network, RAN, side, wherein the selection selects the at least one selected NN (e.g. 40a) based on a signalling from the user equipment, UE, indicating at least one measurement, and further configured to send an indication, in the scheduling, of the selection to be performed.

[0379] The chancel decoder apparatus may receive, from a transmitter, at least one of the measurements on the channel.

[0380] The channel- decoder NN may be trained using a minimization of a loss function and by performing backpropagation towards an original version, at the decoder side, of the source signal.

[0381] The channel decoder apparatus may be trained using the minimization of the loss function during inference.

[0382] The channel decoder NN may include at least one convolutional layer. FHIIS24EM36-2024373152. DOCX 57

[0383] The channel decoder NN may include multiple ID convolutional layers followed by a dense layer.

[0384] The dense layer may be followed by a normalization layer.

[0385] The channel-decoder NN may be trained without minimization of any loss function from a reconstructed version (e.g. 5) of the source signal.

[0386] The channel decoder apparatus may further comprise a different channel decoder unit (e.g. 30b) configured to decode further data (e.g. 5b) which are not provided to the channeldecoder NN.

[0387] The channel decoder apparatus may separate the further data (e.g. lb) from the symbols, or the other representation, representing the source signal (e.g. la), to thereby provide the symbols representing the source signal (e.g. la), or the other representation, to the channel-decoder NN (e.g. 40a), and the further data (e.g. 5b) to the different channel decoder unit (e.g. 40b).

[0388] The channel decoder apparatus may retrieve further data to the symbols or the other representation which is header information, which is not provided to the channel-decoder NN (e.g. 40a).

[0389] The header information may be or may include TCP / IP header or UDP / IP header.

[0390] The header information may include information on length of the symbols or the other representation (e.g. 22a).

[0391] The header information may include information redundancy check information.

[0392] The channel decoder apparatus may be configured to process the non-media data to provide side information on the symbols or the other representation (e.g. 22a).

[0393] The channel decoder apparatus may deencapsulate (e.g. 46) the header information (e.g. 5b, l'b) from the other representation (e.g. 22a) upstream to the source decoder (e.g. 50). FHIIS24EM36-2024373152. DOCX 58

[0394] The different channel decoder unit (e.g. 30b) may decompress the TCP / IP header with using an inverse robust header compression, RoHC.

[0395] The different channel encoder unit (e.g. 30b) may compress the TCP / IP header with robust header compression, RoHC.

[0396] The decoder apparatus may generate the further data (e.g. lb).

[0397] The channel decoder apparatus NN may be trained and may be configured to process the sequence of symbols or the other representation by generating one single vectorial feature representation from one single sequence of symbols.

[0398] The source decoder may apply a neural network to the vectorial feature representation to synthesize the source signal.

[0399] The vectorial feature representation may be either a learned latent representation or a representation of the source signal in a time-frequency domain or in a linear prediction residual domain.

[0400] The channel decoder apparatus may apply a decryption-unit NN (e.g. 42) to decrypt the symbols or the other representation (e.g. 42a), downstream to the channel- decoder NN, the encryption-unit NN receiving the symbols or the other representation and providing a decrypted version of the symbols or the other representation, representing the vectorial feature representation (e.g. 25).

[0401] The channel decoder apparatus may choose between a high-bitrate path (e.g. 41a') and a low-bitrate path (e.g. 41a"), wherein the channel decoder apparatus (e.g. 40) is configured, in the high-bitrate path, to provide, through the at least one channel-decoder NN (e.g. 40b), from the sequence of symbols or the other representation, the vectorial feature representation in the vectorial format specific of the source decoder (e.g. 20) or another feature representation, FHIIS24EM36-2024373152. DOCX 59 wherein the channel decoder apparatus is configured, in the low-bitrate path, to bypass the at least one channel-decoder NN (e.g. 40b), and to thereby provide the vectorial feature representation or the other feature representation in a more compressed form.

[0402] The decoder may further provide the vectorial feature representation or the other feature representation to a compressor (e.g. 140) not using the at least one channel-decoder NN, to obtain a lossy version of the vectorial feature representation or the other feature representation.

[0403] The channel decoder apparatus may choose the compressor based on at least one of the source decoder (e.g. 50).

[0404] The decoder may choose between a high-bitrate path and a low-bitrate path based on signalling.

[0405] The vectorial feature representation or the other feature representation may include redundancy information, and the vectorial feature representation or the other feature representation in the more compressed form includes less redundancy information than the vectorial feature representation or the other feature representation.

[0406] The vectorial feature representation or the other feature representation may include redundancy information, and vectorial feature representation or the other feature representation in the more compressed form (e.g. 120a) includes no redundancy information.

[0407] The channel decoder apparatus may, in the case of choosing the low-bitrate path, bypass the channel-decoder NN.

[0408] The channel decoder apparatus may, in the case of choosing the low-bitrate path, obtain the quantized version of the vectorial feature representation or the other feature representation through a Low-density parity-check, LDPC, decoder. FHIIS24EM36-2024373152. DOCX 60

[0409] The channel decoder apparatus may in the case of choosing the low-bitrate path (e.g. 21a"), obtain the quantized version (e.g. 120a) of the vectorial feature representation (e.g. 20a, 21a) or the other feature representation in the more compressed form through a Polar decoder.

[0410] The channel decoder apparatus may in the case of choosing the low-bitrate path, obtain the quantized version of the vectorial feature representation or the other feature representation in the more compressed form through a Turbo decoder.

[0411] The channel decoder apparatus may decrypt the vectorial feature representation or the other feature representation in the high-bitrate path.

[0412] The channel decoder apparatus may refrain from decrypting the version of the vectorial feature representation or the other feature representation in the more compressed form in the low-bitrate path.

[0413] The channel-decoder NN (e.g. 40a) may retrieve, for at least one symbol or sequence of symbols or at least one value of the other representation, a reliability value associated to the at least one symbol or sequence of symbols or at least one value, so as to provide the reliability value to the source decoder.

[0414] The channel decoder apparatus may read the reliability value from a data field in received side information.

[0415] The channel decoder may measure the reliability value from measurements on the occurrence rate of the at least one symbol or sequence of symbols or at least one value of the other representation.

[0416] The channel decoder apparatus may apply, upstream to the channel-decoder NN (e.g. 40a), a non-NN-based channel-decoder unit (e.g. 40b) which does not decompress the sequence of symbols (e.g. 45) or the other representation of the sequence of symbols (e.g. 45), but performs forward error correction, FEC, information.

[0417] The channel decoder apparatus may read the symbols according to a variable alphabet, to generate the other representation. FHIIS24EM36-2024373152. DOCX 61

[0418] The channel decoder apparatus may read the symbols according to a continuous alphabet, to generate the other representation.

[0419] The channel decoder apparatus may transform the symbols according to a continuous function which maps symbols onto latent values, which generate the other representation.

[0420] The channel decoder apparatus may generate the vectorial latent feature representation and / or the other representation according to a variable number of bits for value.

[0421] The channel decoder apparatus (e.g. 40) may include a frequency selector channel or a channel equalizer upstream to the channel-decoder NN.

[0422] The source signal (e.g. 1) may be an audio signal.

[0423] The source signal (e.g. 1) may be a speech signal.

[0424] The source signal (e.g. 1) may be a text signal.

[0425] The source signal (e.g. 1) may be a video signal.

[0426] The source signal (e.g. 1) may be a media signal.

[0427] The channel decoder apparatus may be implemented in a satellite for satellite-to-satellite communication, satellite-to-earth communication, and / or earth-to-satellite communication.

[0428] A decoder may be comprising the channel decoder of any of the preceding claims and the source decoder.

[0429] The decoder may change among a plurality of source decoders, to thereby select a channel-decoder NN based on the selected source decoder.

[0430] The decoder may have the source encoder at a higher layer than the channel decoder apparatus. FHIIS24EM36-2024373152. DOCX 62

[0431] The decoder may comprise a plurality of source-decoder NNs (e.g. 30a), each of the plurality of source-decoder NNs (e.g. 30a) being associated with at least one of a plurality of source decoders, each of the plurality of the source-decoder NNs (e.g. 30a) being selectable based at least in part on one of the source decoder, the vectorial format specific of the source decoder, a specific operating mode of the source decoder and specific type of source signal.

[0432] In examples above there is provided, inter alia, a channel encoder apparatus (e.g. 30) for encoding a source signal (e.g. 1) received as a vectorial feature representation (e.g. 20a, 21a) from a source encoder (e.g. 20), the vectorial feature representation (e.g. 20a, 21a) being in a vectorial format specific of the source encoder (e.g. 20), the channel encoder apparatus (e.g. 30) being configured to apply at least one channelencoder neural network, NN (e.g. 30a), to the vectorial feature representation (e.g. 20a, 21a) in the vectorial format specific of the source encoder (e.g. 20) or to another feature representation (e.g. 22a, 23a, 24a, 25) of the vectorial feature representation (e.g. 20a, may provide a sequence of symbols (e.g. 35) from the vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25).

[0433] The source encoder may apply a neural network to the source signal to obtain the vectorial feature representation. at least one NN (e.g. 30a, 42) may be a fully-connected NN or is a recurrent NN which, when developed, is a fully connected NN.

[0434] The at least one NN (e.g. 30a, 42) may be such that the NN is not separable in multiple sub-NNs independent from each other.

[0435] The vectorial feature representation may be either a learned feature representation.

[0436] The vectorial feature representation may be a representation of the source signal derived after a time-frequency transformation or a linear prediction.

[0437] The sequence of symbols may be a sequence of modulation symbols. FHIIS24EM36-2024373152. DOCX 63

[0438] The source signal (e.g. 1) may be a media signal.

[0439] The channel-encoder NN (e.g. 30a) may be configured to process the vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25) through a sequence of layers using weights, the weights weighting, in parallel with each other, the vectorial feature representation (e.g. 20a, 21a), or the other feature representation (e.g. 22a, 23a, 24a, 25).

[0440] The channel-encoder NN (e.g. 30a) may be configured to process the vectorial feature representation (e.g. 20a, or the other feature representation (e.g. 22a, 23a, 24a, 25) through a sequence of layers using weights having fixed position bounded to the vectorial format specific of the source encoder (e.g. 20).

[0441] The channel encoder apparatus may perform a selection selecting, among a plurality of selectable channel-encoder NNs, a selected at least one channel-encoder NN based at least in part on the specific source encoder.

[0442] The channel encoder apparatus may perform a selection selecting, among a plurality of selectable channel-encoder NNs, a selected at least one channel-encoder NN based at least in part on the vectorial format specific of the source encoder (e.g. 20).

[0443] The channel encoder apparatus may perform a selection selecting, among a plurality of selectable channel-encoder NNs, a selected at least one channel-encoder NN based at least in part on a specific operating mode of the source encoder.

[0444] The channel encoder apparatus may perform a selection selecting, among a plurality of selectable channel-encoder NNs, a selected at least one channel-encoder NN based at least in part on a specific type of source signal.

[0445] The channel encoder apparatus may perform the selection based on an indication from the source encoder (e.g. 20). FHIIS24EM36-2024373152. DOCX 64

[0446] The channel encoder apparatus may perform the selection based on a request from a receiver.

[0447] The channel encoder apparatus may perform the selection based on a request from a higher layer.

[0448] The channel encoder apparatus may perform the selection based on a request from a physical layer.

[0449] The channel encoder apparatus may receive, from a receiver, a request for indicating whether the channel encoder apparatus has the capability of performing the selection, and may respond that the channel encoder apparatus has the capability of performing the selection.

[0450] The channel encoder apparatus may carry out, with a receiver, a handshaking procedure defining, based at least in part on the request from the receiver, on the source encoder, and on selectable NNs, the at least one selected NN.

[0451] The channel encoder apparatus may receive, from a higher layer, a request for indicating whether the channel encoder apparatus has the capability of performing the selection, and to respond that the channel encoder apparatus has the capability of performing the selection.

[0452] The selection may select the at least one selected NN (e.g. 30a) irrespective of the channel conditions.

[0453] The selection may select the at least one selected NN (e.g. 30a) based, at least partially, on the measurements on the channel.

[0454] The chancel encoder apparatus may perform at least one of the measurements on the channel. FHIIS24EM36-2024373152. DOCX 65

[0455] The channel encoder apparatus may be implemented in a user equipment, UE, the channel encoder being configured to perform the at least one of the measurements through Reference Signal Received Power, RSRP.

[0456] The channel encoder apparatus may be implemented in a user equipment, UE, the channel encoder being configured to perform the at least one of the measurements through Reference Signal Received Quality, RSRQ.

[0457] The channel encoder apparatus may report, to a radio access network, RAN, side, the measurements, subsequently waiting for an indication, in the scheduling, of the selection to be performed.

[0458] The channel decoder apparatus may be implemented at a radio access network, RAN, side, wherein the selection selects the at least one selected NN (e.g. based on a signalling from the user equipment, UE, indicating at least one measurement, and further configured to send an indication, in the scheduling, of the selection to be performed.

[0459] The chancel encoder apparatus may receive, from a receiver, at least one of the measurements on the channel.

[0460] The channel-encoder NN (e.g. 30a) may be trained using a backpropagation and a minimization of a loss function from a reconstructed version (e.g. 45) of the vectorial feature representation (e.g. 20a) or the other feature representation (e.g. 22a, 23a, 24a, 25).

[0461] The channel-encoder NN (e.g. 30a) may be optimized by error minimization of an error between a reconstructed version (e.g. 45) of the vectorial feature representation (e.g. 20a) or the other feature representation (e.g. 22a, 23a, 24a, 25) and an estimated version of the vectorial feature representation (e.g. 20a) or the other feature representation (e.g. 22a, 23a, 24a, 25).

[0462] The channel encoder apparatus may be trained using the minimization of the loss function during inference.

[0463] The channel-decoder NN may include at least one convolutional layer. FHIIS24EM36-2024373152. DOCX 66

[0464] The channel-decoder NN may include multiple ID convolutional layers followed by a dense layer.

[0465] The dense layer may be followed by a normalization layer.

[0466] The channel-encoder NN may be trained without minimization of any loss function from a reconstructed version (e.g. 5) of the source signal.

[0467] The channel encoder apparatus may further comprise a different channel encoder unit (e.g. 30b) configured to encode further data (e.g. lb) which are not provided to the channelencoder NN (e.g. 30a).

[0468] The channel encoder apparatus may separate the further data (e.g. lb) from the source signal (e.g. la), to thereby provide the vectorial feature representation (e.g. 20a, 21a) or the other representation to the channel-encoder NN (e.g. 30a), and the further data (e.g. lb) to the different channel encoder unit (e.g. 30b).

[0469] The channel encoder apparatus may insert (e.g. 24) the further data (e.g. lb), or part of it, to the vectorial feature representation (e.g. 20a, 21a) or the other representation (e.g. 22a), the further data, or part of it, being header information, which is not provided to the channel-encoder NN (e.g. 30a).

[0470] The header information may be or include TCP / IP header or UDP / IP header.

[0471] The header information may include information on length of the vectorial feature representation (e.g. 20a, 21a) or the other representation (e.g. 22a).

[0472] The header information may include information redundancy check information.

[0473] The channel encoder apparatus may process the non-media data to provide side information on the vectorial feature representation (e.g. 20a, 21a) or the other representation (e.g. 22a). FHIIS24EM36-2024373152. DOCX 67

[0474] The channel encoder apparatus may further separate (e.g. 26) the header information from the vectorial feature representation (e.g. 20a, 21a) or the other representation (e.g. 22a) before the channel-encoder NN (e.g. 30a).

[0475] The different channel encoder unit (e.g. 30b) may compress the TCP / IP header with robust header compression, RoHC.

[0476] The encoder apparatus may generate the further data (e.g. lb).

[0477] The channel-encoder NN may be trained and / or may process the vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25) without separating the components of the vectorial feature representation (e.g. 20a, 21a) or of the other feature representation (e.g. 22a, 23a, 24a, 25).

[0478] The channel-encoder NN may be trained and / or may provide the sequence of symbols (e.g. 35) without quantizing the vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25).

[0479] The channel encoder apparatus may apply an encrypting NN (e.g. 22) to encrypt the vectorial feature representation (e.g. 20a, 21a) or the other feature representation upstream to the channel-encoder NN or as a part of the channel-encoder NN, the encrypting NN receiving the vectorial feature representation (e.g. 20a, 21a) and providing an encrypted feature (e.g. 22a) representing the vectorial feature representation (e.g. 20a, 21a).

[0480] The channel encoder apparatus may be trained by inserting random data in addition to the vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25) providing the information on the vectorial feature representation (e.g. 20a, 21a).

[0481] The channel encoder apparatus may be so that the random data has the ergodic properties as the vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25) providing the information on the vectorial feature representation (e.g. 20a, 21a). FHIIS24EM36-2024373152. DOCX 68

[0482] The channel encoder apparatus may choose (e.g. 219 between a high-bitrate path (e.g. 21a7) and a low-bitrate path (e.g. 21a"), the channel encoder apparatus may, in the high-bitrate path (e.g. 21a'), provide, through the at least one channel-encoder NN (e.g. 30a), the sequence of symbols (e.g. 35) from the vectorial feature representation (e.g. 20a, 21a) in the vectorial format specific of the source encoder (e.g. 20) or the other feature representation (e.g. 22a, 23a, 24a, 25), the channel encoder apparatus may, in the low-bitrate path (e.g. 21a"), provide the vectorial feature representation (e.g. 20a, 21a) or the other feature representation to a compressor (e.g. 120), to obtain a lossy-compressed version (e.g. 120a) of the vectorial feature representation (e.g. 20a, 21a) or the other feature representation, so as to encode (e.g. 30a) the lossy-compressed version (e.g. 120a) of the vectorial feature representation (e.g. 20a, 21a) or the other feature representation.

[0483] The vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25) may include redundancy information, and the lossy-compressed version (e.g. 120a) of the vectorial feature representation (e.g. 20a, 21a) or the other feature representation may include less redundancy information than the vectorial feature representation (e.g. 20a, 21a) or the other feature representation.

[0484] The channel encoder apparatus may choose the compressor based on at least one of the source encoder (e.g. 20).

[0485] The vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25) may include redundancy information, and the lossy-compressed version (e.g. 120a) of the vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25)may include no redundancy information.

[0486] The channel encoder apparatus may, in the case of choosing the low-bitrate path (e.g. 21a"), bypass the channel-encoder NN (e.g. 30a).

[0487] The channel encoder apparatus may, in the case of choosing the low-bitrate path (e.g. 21a"), provide the quantized version (e.g. 120a) of the vectorial feature representation (e.g. 20a, 21a) or the other feature representation to a Low-density parity-check, LDPC, encoder. FHIIS24EM36-2024373152. DOCX 69

[0488] The channel encoder apparatus may, in the case of choosing the low-bitrate path (e.g. 21a"), provide the quantized version (e.g. 120a) of the vectorial feature representation (e.g. 20a, 21a) or the other feature representation to a Polar encoder.

[0489] The channel encoder apparatus may, in the case of choosing the low-bitrate path (e.g. 21a"), provide the quantized version (e.g. 120a) of the vectorial feature representation (e.g. 20a, 21a) or the other feature representation to a Turbo encoder.

[0490] The channel encoder apparatus may send a first, high-bitrate version, of the source signal (e.g. 1) towards a first receiver through the high-bitrate path (e.g. 21a7), and a second, low-bitrate version (e.g. 120a), of the source signal (e.g. 1) towards a second receiver through the low-bitrate path (e.g. 21a").

[0491] The channel encoder may send the first, high-bitrate version, of the source signal (e.g. 1) towards the first receiver through the high-bitrate path (e.g. 21a7) as a wireless signal, and the second, low-bitrate version (e.g. 120a), of the source signal (e.g. 1) towards the second receiver through the low-bitrate path (e.g. 21a") as an electric signal.

[0492] The channel encoder apparatus may encrypt (e.g. 22) the vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25) in the high- bitrate path (e.g. 21a7), and may not encrypt quantized version (e.g. 120a) of the vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25) in the low- bitrate path (e.g. 21a").

[0493] The channel encoder apparatus may encrypt (e.g. 22) the vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25) in the high- bitrate path (e.g. 21a7) using a NN (e.g. 22), and may not encrypt quantized version (e.g. 120a) of the vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25) in the low-bitrate path (e.g. 21a"), or may encrypt the quantized version (e.g. 120a) of the vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25) in the low-bitrate path without a NN. FHIIS24EM36-2024373152. DOCX 70

[0494] The channel encoder apparatus may provide the sequence of symbols as a continuous sequence of symbols.

[0495] The channel encoder apparatus may provide the sequence of symbols as a continuous sequence of symbols variable whit the particular channel-decoder NN chosen.

[0496] The channel encoder apparatus may provide the sequence of symbols as modulation symbols.

[0497] The channel-encoder NN (e.g. 30a) may provide, for at least one value outputted by the NN (e.g. 30a), a reliability value associated to the value, so as to encode the probabilistic value in the side information and / or in the sequence of symbols.

[0498] The channel encoder apparatus may generate the symbols according to a variable alphabet.

[0499] The channel encoder apparatus may generate the symbols according to a continuous alphabet.

[0500] The channel encoder apparatus may generate the symbols according to a continuous function which maps latent values onto symbols.

[0501] The channel encoder apparatus may apply, downstream to the channel-encoder NN (e.g. 30a), a non-NN-based channel-encoder unit (e.g. 30b) which inserts forward error correction, FEC, information.

[0502] The channel encoder apparatus may be implemented in a satellite for satellite-to-satellite communication, satellite-to-earth communication, and / or earth-to-satellite communication.

[0503] An encoder for encoding a media signal (e.g. la), may comprise the channel encoder apparatus and the source encoder apparatus, may transmit the sequence of symbols (e.g. 35) from the vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25). FHIIS24EM36-2024373152. DOCX 71

[0504] The source encoder may apply a neural network to the source signal to obtain the vectorial feature representation.

[0505] A user equipment, UE, may comprise the channel encoder apparatus and the source encoder apparatus, wherein the channel encoder apparatus is configured to transmit the sequence of symbols (e.g. 35) from the vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25) in the uplink, the UE being configured to send signalling indicating that it has the channel encoder apparatus (e.g. 30a) with the channel-encoder NN (e.g. 30a).

[0506] A user equipment, UE, may comprise the channel decoder apparatus and the source decoder apparatus, wherein the channel decoder apparatus is configured to generate a vectorial feature representation (e.g. 45), as a vectorial feature representation specific of a source decoder, from a sequence of symbols (e.g. 36), the vectorial representation representing a source signal, in the downlink, the UE being configured to send signalling indicating that it has the channel decoder apparatus (e.g. 40a) with the channel-decoder NN (e.g. 40a).

[0507] The UE may reply to a request from a core network.

[0508] The UE may reply to a request from a radio access network, RAN.

[0509] The UE signalling may be NAS signalling and may be directed to a core network.

[0510] The UE signalling may be on RRC layer.

[0511] The UE signalling may be in 5GS session management.

[0512] The UE signalling may be in control information, DCI, for scheduling of PUSCH.

[0513] A radio access network, RAN (e.g. 300), may comprise the channel encoder apparatus and the source encoder, wherein the channel encoder apparatus is configured to transmit to a user equipment, UE, the sequence of symbols (e.g. 35) from the vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25), FHIIS24EM36-2024373152. DOCX 72 the RAN (e.g. 300) being configured to send signalling (e.g. 302) requesting the UE (e.g. 200) to send information whether the UE has the channel encoder apparatus (e.g. 30a) with the channel-encoder NN (e.g. 30a) or whether the UE has the has the channel decoder apparatus (e.g. 40) with the channel-decoder NN (e.g. 40a).

[0514] The RAN may send the request on behalf of a core network.

[0515] The RAN may , in case of positive response from the UE, start generating or receiving the symbols using the channel-encoder NN (e.g. 30a) of the channel encoder apparatus, and, in case of negative response from the UE, may start generating symbols using a different channel encoder unit (e.g. 30b) which does not use the channel-encoder NN (e.g. 30a).

[0516] The positive or negative response may be NAS signalling and may be directed to the core network.

[0517] The positive or negative response may be on RRC layer.

[0518] The positive or negative response may be in 5GS session management.

[0519] The signalling and / or the positive or negative response may be on physical layer.

[0520] The source signal (e.g. 1) may be a media signal.

[0521] The source signal (e.g. 1) may be an audio signal.

[0522] The source signal (e.g. 1) may be a speech signal.

[0523] The source signal (e.g. 1) may be a text signal.

[0524] The source signal (e.g. 1) may be a video signal.

[0525] An encoder may comprise the channel encoder apparatus and the source encoder. FHIIS24EM36-2024373152. DOCX 73

[0526] The encoder may choose among a plurality of source encoders, to thereby select the channel-encoder NN based on the selected source encoder.

[0527] The encoder may have the source encoder at a higher layer than the channel decoder apparatus.

[0528] The encoder may be comprise a plurality of source-encoder NNs (e.g. 30a), each of the plurality of source-encoder NNs (e.g. 30a) being associated with at least one of a plurality of source encoders, each of the plurality of the source-encoder NNs (e.g. 30a) being selectable based at least in part on one of the source decoder, the vectorial format specific of the source decoder (e.g. 20), a specific operating mode of the source decoder and specific type of source signal.

[0529] In examples above there is provided, inter alia, a method for generating a vectorial feature representation as a vectorial feature representation (e.g. 40a, 45) in a format specific of a source decoder (e.g. 50), from a sequence of symbols (e.g. 36), the vectorial feature representation (e.g. 40a) representing a source signal (e.g. 1, 5), the method including applying at least one channel-decoder neural network, NN (e.g. 40a, 42), to the sequence of symbols (e.g. 36) or another representation (e.g. 46a, 43a, 42a) of the sequence of symbols, thereby generating the vectorial feature representation in the format specific of the source decoder (e.g. 50).

[0530] In examples above there is provided, inter alia, a method for encoding a source signal (e.g. 1) received as a vectorial feature representation (e.g. 20a, 21a) from a source encoder (e.g. 20), the vectorial feature representation (e.g. 20a, 21a) being in a vectorial format specific of the source encoder (e.g. 20), the method including applying at least one channel-encoder neural network, NN (e.g. 30a), to the vectorial feature representation (e.g. 20a, 21a) in the vectorial format specific of the source encoder (e.g. 20) or to another feature representation (e.g. 22a, 23a, 24a, 25) of the vectorial feature representation (e.g. 20a, 21a), to provide a sequence of symbols (e.g. 35) from the vectorial feature representation (e.g. 20a, 21a) or the other feature representation (e.g. 22a, 23a, 24a, 25). FHIIS24EM36-2024373152. DOCX 74

[0531] In examples above there is provided, inter alia, a non-transitory storage unit storing instructions which, when executed by a processor, may cause the processor to perform a method according to the method of examples mentioned above.

Claims

FHIIS24EM36-2024373152. DOCX 75Claims1. A channel decoder apparatus (40) for generating a vectorial feature representation as a vectorial feature representation (40a, 45) in a format specific of a source decoder (50), from a sequence of symbols (36), the vectorial feature representation (40a) representing a source signal (1, 5), the channel decoder apparatus being configured to apply at least one channeldecoder neural network, NN (40a, 42), to the sequence of symbols (36) or another representation (46a, 43a, 42a) of the sequence of symbols, thereby generating the vectorial feature representation in the format specific of the source decoder (50).

2. The channel decoder apparatus of any of the preceding claims wherein the at least one NN (40a, 42) is a fully-connected NN or is a recurrent NN which, when developed, is a fully connected NN.

3. The channel decoder apparatus of any of the preceding claims wherein the at least one NN (40a, 42) is such that the NN is not separable in multiple sub-NNs independent from each other.

4. The channel decoder apparatus of any of the preceding claims, wherein the channeldecoder NN is configured to process the sequence of symbols or the other representation, through a sequence of layers using weights, the weights weighting, in parallel with each other, the sequence of symbols or the other representation.

5. The channel decoder apparatus of any of the preceding claims, configured to perform a selection selecting, among a plurality of selectable channel-decoder NNs, a selected at least one channel-decoder NN based at least in part on the specific source decoder.

6. The channel decoder apparatus of any of claims 2-5, configured to perform a selection selecting, among a plurality of selectable channel- decoder NNs, a selected at least one channel-decoder NN based at least in part on the vectorial format specific of the source decoder (20).FHIIS24EM36-2024373152. DOCX 767. The channel decoder apparatus of any of claims 5-6, configured to perform a selection selecting, among a plurality of selectable channel-decoder NNs, a selected at least one channel-decoder NN based at least in part on a specific operating mode of the source decoder.

8. The channel decoder apparatus of any of claims 5-7, configured to perform a selection selecting, among a plurality of selectable channel- decoder NNs, a selected at least one channel-decoder NN based at least in part on a specific type of source signal.

9. The channel decoder apparatus of any of claims 5-8, configured to perform the selection based on an indication from the source decoder (20).

10. The channel decoder apparatus of any of claims 5-9, configured to perform the selection based on a request from a transmitter.

11. The channel decoder apparatus of any of claims 5-10, configured to perform the selection based on a request from a higher layer.

12. The channel decoder apparatus of any of claims 5-11, configured to receive, from a transmitter, a request for indicating whether the channel decoder apparatus has the capability of performing the selection, and to respond that the channel decoder apparatus has the capability of performing the selection.

13. The channel decoder apparatus of any of claims 5-12, configured to carry out, together with a transmitter, a handshaking procedure defining, based at least in part on the request from the receiver, on the source decoder, and on selectable NNs, the at least one selected NN.

14. The channel decoder apparatus of any of claims 5-13, configured to receive, from a higher layer, a request for indicating whether the channel decoder apparatus has the capability of performing the selection, and to respond that the channel decoder apparatus has the capability of performing the selection.FHIIS24EM36-2024373152. DOCX 7715. The channel decoder apparatus of any of claims 5-14, wherein the selection selects the at least one selected NN (40a) irrespective of the channel conditions.

16. The channel decoder apparatus of any of claims 5-15, wherein the selection selects the at least one selected NN (40a) based, at least partially, on the measurements on the channel.

17. The chancel decoder apparatus of claim 16, configured to perform at least one of the measurements on the channel.

18. The channel decoder apparatus of claim 17, implemented in a user equipment, UE, the channel decoder being configured to perform the at least one of the measurements through Reference Signal Received Power, RSRP.

19. The channel decoder apparatus of any of claims 16-17, implemented in a user equipment, UE, the channel decoder being configured to perform the at least one of the measurements through Reference Signal Received Quality, RSRQ.

20. The channel decoder apparatus of any of claims 16-17, implemented at a user equipment, UE, configured to report, to a radio access network, RAN, side the measurements, subsequently waiting for an indication, in the scheduling, of the selection to be performed.

21. The channel decoder apparatus of any of claims 16-20, implemented at a radio access network, RAN, side, wherein the selection selects the at least one selected NN (40a) based on a signalling from the user equipment, UE, indicating at least one measurement, and further configured to send an indication, in the scheduling, of the selection to be performed.

22. The chancel decoder apparatus of any of claims 16-21, configured to receive, from a transmitter, at least one of the measurements on the channel.FHIIS24EM36-2024373152. DOCX 7823. The channel decoder apparatus of any of the preceding claims, wherein the channeldecoder NN is trained using a minimization of a loss function and by performing backpropagation towards an original version, at the decoder side, of the source signal.

24. The channel decoder apparatus of claim 23, configured to be trained using the minimization of the loss function during inference.

25. The channel decoder apparatus of any of the preceding claims, wherein the channeldecoder NN includes at least one convolutional layer.

26. The channel decoder apparatus of claim 25, wherein the channel-decoder NN includes multiple ID convolutional layers followed by a dense layer.

27. The channel decoder apparatus of claim 26, wherein the dense layer is followed by a normalization layer.

28. The channel decoder apparatus of any of the preceding claims, wherein the channeldecoder NN is trained without minimization of any loss function from a reconstructed version (5) of the source signal.

29. The channel decoder apparatus of any of the preceding claims, further comprising a different channel decoder unit (30b) configured to decode further data (5b) which are not provided to the channel-decoder NN.

30. The channel decoder apparatus of claim 29, configured to separate the further data (lb) from the symbols, or the other representation, representing the source signal (la), to thereby provide the symbols representing the source signal (la), or the other representation, to the channel-decoder NN (40a), and the further data (5b) to the different channel decoder unit (40b).

31. The channel decoder apparatus of claim 30, configured to retrieve further data to the symbols or the other representation which is header information, which is not provided to the channel-decoder NN (40a).FHIIS24EM36-2024373152. DOCX 7932. The decoder apparatus of claim 31, wherein the header information is or includes TCP / IP header or UDP / IP header.

33. The channel decoder apparatus of any of claims 31-32, wherein the header information includes information on length of the symbols or the other representation (22a).

34. The channel decoder apparatus of any of claims 31-33, wherein the header information includes information redundancy check information.

35. The channel decoder apparatus of any of claims 31-34, configured to process the non-media data to provide side information on the symbols or the other representation (22a).

36. The channel decoder apparatus of any of claims 31-35, configured to deencapsulate (46) the header information (5b, l'b) from the other representation (22a) upstream to the source decoder (50).

37. The decoder apparatus of any of claims 29-36, wherein the different channel decoder unit (30b) is configured to decompress the TCP / IP header with using an inverse robust header compression, RoHC.

38. The decoder apparatus of any of the preceding claims , wherein the different channel encoder unit (30b) is configured to compress the TCP / IP header with robust header compression, RoHC.

39. The decoder apparatus of any of the preceding claims, configured to generate the further data (lb).

40. The channel decoder apparatus of any of the preceding claims, wherein the channeldecoder NN is trained and / or is configured to process the sequence of symbols or the other representation by generating one single vectorial feature representation from one single sequence of symbols.FHIIS24EM36-2024373152. DOCX 8041. The channel decoder apparatus of any of the preceding claims wherein the source decoder applies a neural network to the vectorial feature representation to synthesize the source signal.

42. The channel decoder apparatus of any of the preceding claims wherein the vectorial feature representation is either a learned latent representation or a representation of the source signal in a time-frequency domain or in a linear prediction residual domain.

43. The channel decoder apparatus of any of claims, further configured to apply a decryption-unit NN (42) to decrypt the symbols or the other representation (42a), downstream to the channel- decoder NN, the encryption-unit NN receiving the symbols or the other representation and providing a decrypted version of the symbols or the other representation, representing the vectorial feature representation (25).

44. The channel decoder apparatus of any of the preceding claims, configured to choose between a high-bitrate path (41a') and a low-bitrate path (41a"), wherein the channel decoder apparatus (40) is configured, in the high-bitrate path, to provide, through the at least one channel-decoder NN (40b), from the sequence of symbols or the other representation, the vectorial feature representation in the vectorial format specific of the source decoder (20) or another feature representation, wherein the channel decoder apparatus is configured, in the low-bitrate path, to bypass the at least one channel-decoder NN (40b), and to thereby provide the vectorial feature representation or the other feature representation in a more compressed form.

45. The decoder of claim 44, configured to further provide the vectorial feature representation or the other feature representation to a compressor (140) not using the at least one channel-decoder NN, to obtain a lossy version of the vectorial feature representation or the other feature representation.

46. The channel decoder apparatus of any of claims 44-45, configured to choose the compressor based on at least one of the source decoder (50).

47. The decoder of any of claims 44-46, configured to choose between a high-bitrate path and a low-bitrate path based on signalling.FHIIS24EM36-2024373152. DOCX 8148. The decoder apparatus of any of claims 44-47, wherein the vectorial feature representation or the other feature representation includes redundancy information, and the vectorial feature representation or the other feature representation in the more compressed form includes less redundancy information than the vectorial feature representation or the other feature representation.

49. The channel decoder apparatus of any of claims 44-47, wherein the vectorial feature representation or the other feature representation includes redundancy information, and vectorial feature representation or the other feature representation in the more compressed form (120a) includes no redundancy information.

50. The channel decoder apparatus of any of claims 44-49, configured, in the case of choosing the low-bitrate path, to bypass the channel-decoder NN.

51. The channel decoder apparatus of any of claims 44-50, configured, in the case of choosing the low-bitrate path, to obtain the quantized version of the vectorial feature representation or the other feature representation through a Low-density parity-check, LDPC, decoder.

52. The channel decoder apparatus of any of claims 44-50, configured, in the case of choosing the low-bitrate path (21a"), to obtain the quantized version (120a) of the vectorial feature representation (20a, 21a) or the other feature representation in the more compressed form through a Polar decoder.

53. The channel decoder apparatus of any of claims 44-50, configured, in the case of choosing the low-bitrate path, to obtain the quantized version of the vectorial feature representation or the other feature representation in the more compressed form through a Turbo decoder.

54. The channel decoder apparatus of any claims 44-53, configured to decrypt the vectorial feature representation or the other feature representation in the high-bitrate path.FHIIS24EM36-2024373152. DOCX 8255. The channel decoder apparatus of claim 54, configured to refrain from decrypting the version of the vectorial feature representation or the other feature representation in the more compressed form in the low-bitrate path.

56. The channel decoder apparatus of any of the preceding claims, wherein the channeldecoder NN (40a) is configured to retrieve, for at least one symbol or sequence of symbols or at least one value of the other representation, a reliability value associated to the at least one symbol or sequence of symbols or at least one value, so as to provide the reliability value to the source decoder.

57. The channel decoder apparatus of claim 56, configured to read the reliability value from a data field in received side information.

58. The channel decoder apparatus of claim 56, configured to measure the reliability value from measurements on the occurrence rate of the at least one symbol or sequence of symbols or at least one value of the other representation.

59. The channel decoder apparatus of any of the preceding claims, configured to apply, upstream to the channel-decoder NN (40a), a non-NN-based channel-decoder unit (40b) which does not decompress the sequence of symbols (45) or the other representation of the sequence of symbols (45), but performs forward error correction, FEC, information.

60. The channel decoder apparatus of any of the preceding claims, configured to read the symbols according to a variable alphabet, to generate the other representation.61 The channel decoder apparatus of any of the preceding claims, configured to read the symbols according to a continuous alphabet, to generate the other representation.

62. The channel decoder apparatus of any of the preceding claims, configured to transform the symbols according to a continuous function which maps symbols onto latent values, which generate the other representation.FHIIS24EM36-2024373152. DOCX 8363. The channel decoder apparatus of any of the preceding clams, configured to generate the vectorial latent feature representation and / or the other representation according to a variable number of bits for value.

64. The channel decoder apparatus (40) according to any of the preceding claims, including a frequency selector channel or a channel equalizer upstream to the channeldecoder NN.

65. The channel decoder apparatus of any of the preceding claims, wherein the source signal (1) is an audio signal.

66. The channel decoder apparatus of any of the preceding claims, wherein the source signal (1) is a speech signal.

67. The channel decoder apparatus of any of the preceding claims, wherein the source signal (1) is a text signal.

68. The channel decoder apparatus of any of claims 1-64, wherein the source signal (1) is a video signal.

69. The channel decoder apparatus of any of the preceding claims, wherein the source signal (1) is a media signal.

70. The channel decoder apparatus of any of the preceding claims, configured to be implemented in a satellite for satellite-to-satellite communication, satellite-to-earth communication, and / or earth-to-satellite communication.

71. A decoder comprising the channel decoder of any of the preceding claims and the source decoder.

72. The decoder of claim 71, configured to change among a plurality of source decoders, to thereby select a channel-decoder NN based on the selected source decoder.FHIIS24EM36-2024373152. DOCX 8473. The decoder of claim 71 or 72, configured to have the source encoder at a higher layer than the channel decoder apparatus.

74. The decoder of any of claims 71-73, comprising a plurality of source-decoder NNs (30a), each of the plurality of source-decoder NNs (30a) being associated with at least one of a plurality of source decoders, each of the plurality of the source-decoder NNs (30a) being selectable based at least in part on one of the source decoder, the vectorial format specific of the source decoder, a specific operating mode of the source decoder and specific type of source signal.

75. A channel encoder apparatus (30) for encoding a source signal (1) received as a vectorial feature representation (20a, 21a) from a source encoder (20), the vectorial feature representation (20a, 21a) being in a vectorial format specific of the source encoder (20), the channel encoder apparatus (30) being configured to apply at least one channelencoder neural network, NN (30a), to the vectorial feature representation (20a, 21a) in the vectorial format specific of the source encoder (20) or to another feature representation (22a, 23a, 24a, 25) of the vectorial feature representation (20a, 21a), to provide a sequence of symbols (35) from the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25).

76. The channel encoder apparatus of claim 75 wherein the source encoder applies a neural network to the source signal to obtain the vectorial feature representation.

77. The channel encoder apparatus of any of claims 75-76 wherein the at least one NN (30a, 42) is a fully-connected NN or is a recurrent NN which, when developed, is a fully connected NN.

78. The channel encoder apparatus of any of claims 75-77 wherein the at least one NN (30a, 42) is such that the NN is not separable in multiple sub-NNs independent from each other.

79. The channel encoder apparatus of any of the claims 75-78 wherein the vectorial feature representation is either a learned feature representation.FHIIS24EM36-2024373152. DOCX 8580. The channel decoder apparatus of claim 75, wherein the vectorial feature representation is a representation of the source signal derived after a time-frequency transformation or a linear prediction.

81. The channel encoder apparatus of any of the preceding claims, wherein the sequence of symbols is a sequence of modulation symbols.

82. The channel encoder apparatus of any of claims 75-81, wherein the source signal (1) is a media signal.83 The channel encoder apparatus of any of claims 75-82, wherein the channel-encoder NN (30a) is configured to process the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25) through a sequence of layers using weights, the weights weighting, in parallel with each other, the vectorial feature representation (20a, 21a), or the other feature representation (22a, 23a, 24a, 25).

84. The channel encoder apparatus of any of claims 75-83, wherein the channel-encoder NN (30a) is configured to process the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25) through a sequence of layers using weights having fixed position bounded to the vectorial format specific of the source encoder (20).

85. The channel encoder apparatus of any of claims 75-84, configured to perform a selection selecting, among a plurality of selectable channel-encoder NNs, a selected at least one channel-encoder NN based at least in part on the specific source encoder.

86. The channel encoder apparatus of any of claims 75-85, configured to perform a selection selecting, among a plurality of selectable channel-encoder NNs, a selected at least one channel-encoder NN based at least in part on the vectorial format specific of the source encoder (20).

87. The channel encoder apparatus of any of claims 75-86, configured to perform a selection selecting, among a plurality of selectable channel-encoder NNs, a selected at leastFHIIS24EM36-2024373152. DOCX 86 one channel -encoder NN based at least in part on a specific operating mode of the source encoder.

88. The channel encoder apparatus of any of claims 75-87, configured to perform a selection selecting, among a plurality of selectable channel-encoder NNs, a selected at least one channel-encoder NN based at least in part on a specific type of source signal.

89. The channel encoder apparatus of any of claims 85-88, configured to perform the selection based on an indication from the source encoder (20).

90. The channel encoder apparatus of any of claims 85-89, configured to perform the selection based on a request from a receiver.

91. The channel encoder apparatus of any of claims 85-90, configured to perform the selection based on a request from a higher layer.

92. The channel encoder apparatus of any of any of claims 85-91, configured to perform the selection based on a request from a physical layer.

93. The channel encoder apparatus of any of claims 85-92, configured to receive, from a receiver, a request for indicating whether the channel encoder apparatus has the capability of performing the selection, and to respond that the channel encoder apparatus has the capability of performing the selection.

94. The channel encoder apparatus of any of claims 85-93, configured to carry out, with a receiver, a handshaking procedure defining, based at least in part on the request from the receiver, on the source encoder, and on selectable NNs, the at least one selected NN.

95. The channel encoder apparatus of any of claims 85-94, configured to receive, from a higher layer, a request for indicating whether the channel encoder apparatus has the capability of performing the selection, and to respond that the channel encoder apparatus has the capability of performing the selection.FHIIS24EM36-2024373152. DOCX 8796. The chancel encoder apparatus of any of claims 85-95, wherein the selection selects the at least one selected NN (30a) irrespective of the channel conditions.

97. The chancel encoder apparatus of any of claims 85-95, wherein the selection selects the at least one selected NN (30a) based, at least partially, on the measurements on the channel.

98. The chancel encoder apparatus of claim 97, configured to perform at least one of the measurements on the channel.

99. The channel encoder apparatus of claim 98, implemented in a user equipment, UE, the channel encoder being configured to perform the at least one of the measurements through Reference Signal Received Power, RSRP.

100. The channel encoder apparatus of claim 99 implemented in a user equipment, UE, the channel encoder being configured to perform the at least one of the measurements through Reference Signal Received Quality, RSRQ.

101. The channel encoder apparatus of any of claims 99-100, configured to report, to a radio access network, RAN, side, the measurements, subsequently waiting for an indication, in the scheduling, of the selection to be performed.

102. The channel decoder apparatus of claim 97 or 98, implemented at a radio access network, RAN, side, wherein the selection selects the at least one selected NN (30a) based on a signalling from the user equipment, UE, indicating at least one measurement, and further configured to send an indication, in the scheduling, of the selection to be performed.

103. The chancel encoder apparatus of any of claims 97-102, configured to receive, from a receiver, at least one of the measurements on the channel.

104. The channel encoder apparatus of any of claims 97-103, wherein the channelencoder NN (30a) is trained using a backpropagation and a minimization of a loss function from a reconstructed version (45) of the vectorial feature representation (20a) or the other feature representation (22a, 23a, 24a, 25).FHIIS24EM36-2024373152. DOCX 88105. The channel encoder apparatus of any of claims 97-104, wherein the channelencoder NN (30a) is optimized by error minimization of an error between a reconstructed version (45) of the vectorial feature representation (20a) or the other feature representation (22a, 23a, 24a, 25) and an estimated version of the vectorial feature representation (20a) or the other feature representation (22a, 23a, 24a, 25).

106. The channel encoder apparatus of claim 104 or 105, configured to be trained using the minimization of the loss function during inference.

107. The channel encoder apparatus of claims 75-106, wherein the channel-decoder NN includes at least one convolutional layer.

108. The channel encoder apparatus of claim 107, wherein the channel-decoder NN includes multiple ID convolutional layers followed by a dense layer.

109. The channel encoder apparatus of claim 107 or 108, wherein the dense layer is followed by a normalization layer.

110. The channel encoder apparatus of any of claims 75-109, wherein the channelencoder NN is trained without minimization of any loss function from a reconstructed version (5) of the source signal.

111. The channel encoder apparatus of any of claims 75-110, further comprising a different channel encoder unit (30b) configured to encode further data (lb) which are not provided to the channel-encoder NN (30a).

112. The channel encoder apparatus of claim 111, configured to separate the further data (lb) from the source signal (la), to thereby provide the vectorial feature representation (20a, 21a) or the other representation to the channel-encoder NN (30a), and the further data (lb) to the different channel encoder unit (30b).

113. The channel encoder apparatus of claim 112, configured to insert (24) the further data (lb), or part of it, to the vectorial feature representation (20a, 21a) or the otherFHIIS24EM36-2024373152. DOCX 89 representation (22a), the further data, or part of it, being header information, which is not provided to the channel-encoder NN (30a).

114. The channel encoder apparatus of claim 113, wherein the header information is or includes TCP / IP header or LIDP / IP header.

115. The channel encoder apparatus of any of claims 113-114, wherein the header information includes information on length of the vectorial feature representation (20a, 21a) or the other representation (22a).

116. The channel encoder apparatus of any of claims 113-115, wherein the header information includes information redundancy check information.

117. The channel encoder apparatus of any of claims 113-116, configured to process the non-media data to provide side information on the vectorial feature representation (20a, 21a) or the other representation (22a).

118. The channel encoder apparatus of any of claims 113-117, configured to further separate (26) the header information from the vectorial feature representation (20a, 21a) or the other representation (22a) before the channel-encoder NN (30a).

119. The encoder apparatus of any of claims 111-118, wherein the different channel encoder unit (30b) is configured to compress the TCP / IP header with robust header compression, RoHC.

120. The encoder apparatus of any of claims 111-119, configured to generate the further data (lb).

121. The channel encoder apparatus of any of claims 75-120, wherein the channelencoder NN is trained and / or is configured to process the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25) without separating the components of the vectorial feature representation (20a, 21a) or of the other feature representation (22a, 23a, 24a, 25).FHIIS24EM36-2024373152. DOCX 90122. The channel encoder apparatus of any of claims 75-121, wherein the channelencoder NN is trained and / or is configured to provide the sequence of symbols (35) without quantizing the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25).

123. The channel encoder apparatus of any of claims 75-121, further configured to apply an encrypting NN (22) to encrypt the vectorial feature representation (20a, 21a) or the other feature representation upstream to the channel-encoder NN or as a part of the channel-encoder NN, the encrypting NN receiving the vectorial feature representation (20a, 21a) and providing an encrypted feature (22a) representing the vectorial feature representation (20a, 21a).

124. The channel encoder apparatus of claim 123, trained by inserting random data in addition to the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25) providing the information on the vectorial feature representation (20a, 21a).

125. The channel encoder apparatus of claim 123, configured so that the random data has the ergodic properties as the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25) providing the information on the vectorial feature representation (20a, 21a).

126. The channel encoder apparatus of any of claims 75-125, configured to choose (219 between a high-bitrate path (21a') and a low-bitrate path (21a"), wherein the channel encoder apparatus is configured, in the high-bitrate path (21a'), to provide, through the at least one channel-encoder NN (30a), the sequence of symbols (35) from the vectorial feature representation (20a, 21a) in the vectorial format specific of the source encoder (20) or the other feature representation (22a, 23a, 24a, 25), wherein the channel encoder apparatus is configured, in the low-bitrate path (21a"), to provide the vectorial feature representation (20a, 21a) or the other feature representation to a compressor (120), to obtain a lossy-compressed version (120a) of the vectorial feature representation (20a, 21a) or the other feature representation, so as to encode (30a) the lossy-compressed version (120a) of the vectorial feature representation (20a, 21a) or the other feature representation.FHIIS24EM36-2024373152. DOCX 91127. The channel encoder apparatus of claim 126, wherein the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25) includes redundancy information, and the lossy-compressed version (120a) of the vectorial feature representation (20a, 21a) or the other feature representation includes less redundancy information than the vectorial feature representation (20a, 21a) or the other feature representation.

128. The channel encoder apparatus of any of claims 126-127, configured to choose the compressor based on at least one of the source encoder (20).

129. The channel encoder apparatus of any of claims 126-128, wherein the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25) includes redundancy information, and the lossy-compressed version (120a) of the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25) includes no redundancy information.

130. The channel encoder apparatus of any of claims 126-129, configured, in the case of choosing the low-bitrate path (21a"), to bypass the channel-encoder NN (30a).

131. The channel encoder apparatus of any of claims 126-130, configured, in the case of choosing the low-bitrate path (21a"), to provide the quantized version (120a) of the vectorial feature representation (20a, 21a) or the other feature representation to a Low- density parity-check, LDPC, encoder.

132. The channel encoder apparatus of any of claims 126-130, configured, in the case of choosing the low-bitrate path (21a"), to provide the quantized version (120a) of the vectorial feature representation (20a, 21a) or the other feature representation to a Polar encoder.

133. The channel encoder apparatus of any of claims 126-130, configured, in the case of choosing the low-bitrate path (21a"), to provide the quantized version (120a) of the vectorial feature representation (20a, 21a) or the other feature representation to a Turbo encoder.FHIIS24EM36-2024373152. DOCX 92134. The channel encoder apparatus of any of claims 126-130, configured to send a first, high-bitrate version, of the source signal (1) towards a first receiver through the high-bitrate path (21a'), and a second, low-bitrate version (120a), of the source signal (1) towards a second receiver through the low-bitrate path (21a").

135. The channel encoder apparatus of claim 134, configured to send the first, high- bitrate version, of the source signal (1) towards the first receiver through the high-bitrate path (21a') as a wireless signal, and the second, low-bitrate version (120a), of the source signal (1) towards the second receiver through the low-bitrate path (21a") as an electric signal.

136. The channel encoder apparatus of any of claims 126-135, configured to encrypt (22) the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25) in the high-bitrate path (21a'), and to not encrypt quantized version (120a) of the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25) in the low-bitrate path (21a").

137. The channel encoder apparatus of any of claims 126-136, configured to encrypt (22) the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25) in the high-bitrate path (21a') using a NN (22), and to not encrypt quantized version (120a) of the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25) in the low-bitrate path (21a"), or to encrypt the quantized version (120a) of the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25) in the low-bitrate path without a NN.

138. The channel encoder apparatus of any of claims 75-137, configured to provide the sequence of symbols as a continuous sequence of symbols.

139. The channel encoder apparatus of claim 138, configured to provide the sequence of symbols as a continuous sequence of symbols variable whit the particular channel-decoder NN chosen.FHIIS24EM36-2024373152. DOCX 93140. The channel encoder apparatus of any of claims 75-139, configured to provide the sequence of symbols as modulation symbols.

141. The channel encoder apparatus of any of claims 75-140, wherein the channelencoder NN (30a) is configured to provide, for at least one value outputted by the NN (30a), a reliability value associated to the value, so as to encode the probabilistic value in the side information and / or in the sequence of symbols.

142. The channel encoder apparatus of any of claims 75-141., configured to generate the symbols according to a variable alphabet.

143. The channel encoder apparatus of any of claims 75-142, configured to generate the symbols according to a continuous alphabet.

144. The channel encoder apparatus of any of claims 75-143, configured to generate the symbols according to a continuous function which maps latent values onto symbols.

145. The channel encoder apparatus of any of claims 75-144, configured to apply, downstream to the channel-encoder NN (30a), a non-NN-based channel-encoder unit (30b) which inserts forward error correction, FEC, information.

146. The channel encoder apparatus of any of claims 75-145, configured to be implemented in a satellite for satellite-to-satellite communication, satellite-to-earth communication, and / or earth-to-satellite communication.

147. An encoder for encoding a media signal (la), comprising the channel encoder apparatus of any of claims 75-146 and the source encoder apparatus, wherein the channel encoder apparatus is configured to transmit the sequence of symbols (35) from the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25).

148. The media encoder of claim 147 wherein the source encoder is configured to apply a neural network to the source signal to obtain the vectorial feature representation.FHIIS24EM36-2024373152. DOCX 94149. A user equipment, UE, comprising the channel encoder apparatus of any of claims 75-148 and the source encoder apparatus, wherein the channel encoder apparatus is configured to transmit the sequence of symbols (35) from the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25) in the uplink, the UE being configured to send signalling indicating that it has the channel encoder apparatus (30a) with the channel-encoder NN (30a).

150. A user equipment, UE, comprising the channel decoder apparatus of any of claims 1-149 and the source decoder apparatus, wherein the channel decoder apparatus is configured to generate a vectorial feature representation (45), as a vectorial feature representation specific of a source decoder, from a sequence of symbols (36), the vectorial representation representing a source signal, in the downlink, the UE being configured to send signalling indicating that it has the channel decoder apparatus (40a) with the channel-decoder NN (40a).

151. The UE of any of claims 149-150, configured to reply to a request from a core network.152 The UE of any of claims 149-151, configured to reply to a request from a radio access network, RAN.

153. The UE of any of claims 149-152, wherein the signalling is NAS signalling and is directed to a core network.

154. The UE of any of claims 149-153, wherein the signalling is on RRC layer.

155. The UE of any of claims 149-154, wherein the signalling is in 5GS session management.

156. The UE of any of claims 149-155, wherein the signalling is in control information, DCI, for scheduling of PUSCH.FHIIS24EM36-2024373152. DOCX 95157. A radio access network, RAN (300), comprising the channel encoder apparatus of any of claims 75-156 and the source encoder, wherein the channel encoder apparatus is configured to transmit to a user equipment, UE, the sequence of symbols (35) from the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25), the RAN (300) being configured to send signalling (302) requesting the UE (200) to send information whether the UE has the channel encoder apparatus (30a) with the channel-encoder NN (30a) or whether the UE has the has the channel decoder apparatus (40) with the channel-decoder NN (40a).

158. The RAN of any of claims 157-157, configured to send the request on behalf of a core network.

159. The RAN of any of claims 157-158, configured, in case of positive response from the UE, to start generating or receiving the symbols using the channel-encoder NN (30a) of the channel encoder apparatus, and, in case of negative response from the UE, to start generating symbols using a different channel encoder unit (30b) which does not use the channel-encoder NN (30a).160 The RAN of any of claims 157-159, wherein the positive or negative response is NAS signalling and is directed to the core network.

161. The RAN of any of claims 157-160, wherein the positive or negative response is on RRC layer.

162. The RAN of any of claims 157-161, wherein the positive or negative response is in 5GS session management.

163. The RAN of any of claims 157-162, wherein the signalling and / or the positive or negative response is on physical layer.

164. The channel encoder apparatus of any of claims 75-163, wherein the source signal (1) is a media signal.FHIIS24EM36-2024373152. DOCX 96165. The channel encoder apparatus of any of claims 75-164, wherein the source signal (1) is an audio signal.

166. The channel encoder apparatus of any of claims 75-165, wherein the source signal (1) is a speech signal.

167. The channel encoder apparatus of any of claims 75-166, wherein the source signal (1) is a text signal.

168. The channel encoder apparatus of any of claims 75-167, wherein the source signal (1) is a video signal.

169. An encoder comprising the encoder of any of claims 75-168, comprising the channel encoder apparatus and the source encoder.

170. The encoder of claim 169, configured to choose among a plurality of source encoders, to thereby select the channel-encoder NN based on the selected source encoder.

171. The encoder of claim 170, configured to have the source encoder at a higher layer than the channel decoder apparatus.

172. The encoder of any of claims 169-171, comprising a plurality of source-encoder NNs (30a), each of the plurality of source-encoder NNs (30a) being associated with at least one of a plurality of source encoders, each of the plurality of the source-encoder NNs (30a) being selectable based at least in part on one of the source decoder, the vectorial format specific of the source decoder (20), a specific operating mode of the source decoder and specific type of source signal.

173. A method for generating a vectorial feature representation as a vectorial feature representation (40a, 45) in a format specific of a source decoder (50), from a sequence of symbols (36), the vectorial feature representation (40a) representing a source signal (1, 5), the method including applying at least one channel-decoder neural network, NN (40a, 42), to the sequence of symbols (36) or another representation (46a, 43a, 42a) ofFHIIS24EM36-2024373152. DOCX 97 the sequence of symbols, thereby generating the vectorial feature representation in the format specific of the source decoder (50).

174. A method for encoding a source signal (1) received as a vectorial feature representation (20a, 21a) from a source encoder (20), the vectorial feature representation (20a, 21a) being in a vectorial format specific of the source encoder (20), the method including applying at least one channel-encoder neural network, NN (30a), to the vectorial feature representation (20a, 21a) in the vectorial format specific of the source encoder (20) or to another feature representation (22a, 23a, 24a, 25) of the vectorial feature representation (20a, 21a), to provide a sequence of symbols (35) from the vectorial feature representation (20a, 21a) or the other feature representation (22a, 23a, 24a, 25).

175. A non-transitory storage unit storing instructions which, when executed by a processor, cause the processor to perform a method according to any of claims 173-174.

176. The decoder according to any of claim 71-74, wherein the source decoder has constant parameters.

177. The encoder of any of claim 169-172 wherein the source encoder has constant parameters.

Images