APARELHO E MÉTODO PARA PROCESSAR UM SINAL DE INFORMAÇÃO E MEIO LEGÍVEL POR COMPUTADOR

Abstract

An apparatus for processing an information signal, comprises: a feature extractor (100) for extracting a set of features from the information signal, wherein the feature extractor (100) comprises: a raw feature calculator (110) for calculating raw feature results, each raw feature result having at least two raw feature components; and a raw feature compressor (120) for performing a compression of a dynamic range to the at least two raw feature components to obtain at least two compressed raw feature components for each raw feature result, wherein the set of features comprises the compressed raw feature components; and a signal processor (600) for processing the set of features to obtain the processed information signal, wherein the information signal comprises an audio signal, an image signal, or a radar signal.

Description

1 / 66 APPARATUS AND METHOD FOR PROCESSING A COMPUTER-READABLE INFORMATION SIGNAL Descriptive report

[001] The present invention relates to information signal processing and, specifically, to information signal processing using neural networks.

[002] Deep learning (DL)-based solutions to various problems in the field of audio processing have recently become commonplace, demonstrating improved performance over classical signal processing-based methods. However, one of the main issues with most of the methods developed is their complexity, both computational and memory-related, which makes them unstable for development in embedded devices that generally have limited memory, in addition to computational constraints.

[003] A typical processing sequence for DNN-based audio processing methods is illustrated in Figure 2. The input to such methods is usually an audio waveform, also referred to as a time-domain audio signal. The first block in this sequence is the input preprocessing block (block P1). This block basically consists of processing methods to extract features from the audio waveform and perform some form of dimensional alignment to conform to the dimension requirements of the DNN model. Feature extraction in this block can be a deterministic time-frequency transformation, such as a short-term Fourier transform (STFT), or it can be a set of features learned through backpropagation for a specific application. Following the feature extraction step, a multidimensional tensor (typically 2D or 3D) is obtained and then provided as an input to the block. Petition 870250093764, dated 10 / 14 / 2025, page 79 / 154 2 / 66 of core DNN processing (block D1).

[004] Block D1 is the main DNN processing block that operates on the given input to perform the task for which it was designed. For example, in the case of mask-based noise reduction / speech enhancement [ref], D1 might be a DNN mask estimator that estimates a mask to obtain an enhanced version of the audio input feature representation that consists only of the feature representation of the speech component and eliminates the non-speech components.

[005] The output of D1 is then provided as input to the output postprocessor (block P2). The output postprocessing block basically consists of the final processing operations to obtain the final feature representation of the output audio signal and the inverse operation of the feature transformation / extraction applied in P1 to obtain the final audio output. For example, considering the same noise reduction / speech enhancement application situation as above, block P2 would consist of the masking operation, where the mask estimated from D1 is applied to the original feature representation of the input signal, to obtain the enhanced feature representation to which the final inversion operation is applied in order to obtain the output audio waveform.

[006] The input preprocessing block P1 is also indicated with the numerical reference 10, the DNN processing block D1 is also indicated with the numerical reference 20, and the output postprocessing block P2 is indicated by the numerical reference 30.

[007] The learning procedure of a deep learning system can be negatively affected by a very high dynamic range of the input resources. For example, in DNN-based audio processing systems, this effect Petition 870250093764, dated 10 / 14 / 2025, pp. 80 / 154 3 / 66 can be observed in the form of high and low sounds occurring together in an audio signal. In DNN-based image processing systems, for example, this effect can be observed in the form of light and dark regions in the image occurring together in the same image.

[008] In the literature, several input signal normalization techniques / forms have been applied in order to mitigate the issue. A specific input value compression method, known as power law compression, has also been used. In general, power law compression can be applied to any type of input signals (e.g., audio signals, image signals, radar signals).

[009] Next, a general signal model is considered to represent arbitrary types of input signals provided by S = Sre + J Sim = Sabs6'jsPh Where s is a vector containing a number of k input signals that can be expressed, for example, in terms of signal components, real and imaginary, sre and sim, respectively, or in terms of magnitude and phase components, sabs and sph, respectively. For example, when representing an audio signal from one or multiple microphones in the time and frequency domain, s contains K elements corresponding to K microphones, and the real and imaginary components (or the magnitude and phase components) are the real and imaginary components (or magnitude and phase components) that result from the time and frequency transformation (e.g., STFT) of the microphone signals. As a further example, when representing images, the vector s may contain 3 elements per image pixel, representing, for example, RGB values. In this case, the imaginary components (or the phase components) may be equal to zero unless the image is transformed, by Petition 870250093764, dated 10 / 14 / 2025, page 81 / 154 4 / 66 example, with a Fourier transform. With this signal model, the element k can be represented as: Sk = Sk,re + jSk,im = Sk,abse-jSk,ph

[0010] In existing networks, power-law compression is basically applied, for example, only to the magnitude spectrogram of the computed frequency features, that is, Ck = |Sk|a= |Sk,re + jSk,im|a= Sk,absa Where Sk is the k element of s, α is the power-law factor, and Ck is the compressed signal corresponding to the input signal k. Typically, the compressed signal vector c with its Ck elements is subsequently processed using a DNN, resulting in the compressed signal vector d with the Dk elements. After that, just before the final feature inversion step, the power-law compression is reversed to obtain the final representation of the output feature, given by the vector t, that is, Tk = |Dk|1 / a Where Tk is the k element of the final output feature vector t. Note that in the equations above, only the magnitude of the input (or output) signals is subject to compression (and expansion), and the power-law factor is constant.

[0011] In this work, power-law compression is used to compress / limit the dynamic range of each separate component of the input feature values, and the goal is to improve the learning and generalization capabilities of the DNN; that is, power-law compression is applied to the complete input feature representation instead of just a part of the input, such as the magnitude component. Expressed mathematically, the signal compressed with the proposed power-law compression can be expressed, for example, as: Ck = Ck,re + jCk,im = Ck,abse-jCk,ph Petition 870250093764, dated 10 / 14 / 2025, page 82 / 154 5 / 66 Where Ck,re is the compressed real part component, Ck,im is the compressed imaginary part component, Ck,abs is the compressed magnitude component, and Ck,ph is the compressed phase component of the input signal k, respectively.

[0012] Generally, although neural network applications have more or less distribution in the field, the complexity of neural network processing procedures nevertheless plays a significant role not only in embedded systems and specific hardware systems, but also in other applications that do not run on embedded systems. The reason is that neural network complexity results in high processing resource requirements, high power consumption, and associated problems such as processor resource provision, electronic processor cooling requirements, and timing requirements, particularly in the context of limited computational resources.

[0013] It is an object of the present invention to provide an improved concept for processing information signals.

[0014] This objective is achieved by an information signal processing apparatus according to claim 1, an information signal processing method according to claim 15, or a computer program according to claim 16, wherein preferred embodiments are claimed in the dependent claims.

[0015] According to the first aspect of the present invention, an apparatus for processing an information signal, such as an audio signal or an image signal, or a radar signal, comprises a feature extractor for extracting a set of signal features, the feature set having a first dimension. The apparatus further comprises a feature segmenter for segmenting the feature set into a first dimension. Petition 870250093764, dated 10 / 14 / 2025, page 83 / 154 6 / 66 resource subset and a second resource subset, the first resource subset having a second dimension and the second resource subset having a third dimension, wherein the second and third dimensions are less than the first dimension. Additionally, the first and second resource subsets have an overlapping range, such that the first and second resources of the resource set are included in both the first and second subsets.

[0016] Additionally, the apparatus comprises a neural network processor configured to process the first subset using a first neural network in order to obtain a first result and to process the second subset of features using a second neural network in order to obtain a second result. The results are combined by means of a feature combiner in order to combine the first result and the second result using a third neural network, wherein the third neural network has a third complexity that is less than the first complexity of the first neural network, or less than the second complexity of the second neural network, and the output of the feature combiner is a resulting feature set that has a resulting dimension greater than that of the second dimension or the third dimension and, preferably, equal to that of the first dimension.Additionally, an output post-processor is provided to post-process the resulting feature set in order to obtain a processed information signal.

[0017] According to the first aspect of the present invention, low-complexity processing blocks for deep neural networks, based on audio processing methods (DNN), are obtained. In preferred embodiments, the focus is placed on computational complexity in the primary case, that is, more focus is placed on computational complexity than on complexity of Petition 870250093764, dated 10 / 14 / 2025, page 84 / 154 7 / 66 memory. Additionally, the apparatus and the computer methods and programs are, in a preferred embodiment, presented in the context of a neural network-based noise reduction solution. However, the inventive procedures, according to the first aspect, but also according to the second and third aspects to be described later, are also applicable to other neural networks or deep neural network-based solutions for audio problems, or other information signal processing, which require reconstructed information signals at the output. Such applications may, for audio signal applications, comprise dereverberation processing, echo cancellation processing, speech or audio coding applications, bandwidth extension applications, etc.For image signal processing, applications such as image enhancement, edge enhancement, image recognition, or other image processing applications can be improved with the inventive procedures, according to the first, second, and / or third aspects. Additionally, the analysis of radar signals for the purpose of object detection and location also relies on the analysis of features in different signal frequency bands, for example.

[0018] Generally, the complexity or computational complexity of a neural network can be quantified in many different quantities, such as floating-point operations (FLOPS), where a higher number of floating-point operations represents greater complexity, or execution time on one or more target hardware devices, where longer execution time represents greater complexity, or the power consumption of a given device, where higher power consumption represents greater complexity, or the number of MAC operations. Petition 870250093764, dated 10 / 14 / 2025, page 85 / 154 8 / 66 (Multiply and Accumulate), where a larger number of MAC operations represents greater complexity.

[0019] According to the second aspect of the invention, the form in which the extracted features are presented to the signal processor, comprising, by way of illustration, one or more neural networks, is the focus. An apparatus for processing an information signal, according to a second aspect of the invention, comprises a feature extractor for extracting a set of features from the information signal. The feature extractor comprises a raw feature calculator for calculating the raw feature results, wherein each raw feature result has at least two raw feature components. Preferred raw feature components are magnitude and phase on the one hand, or real part and imaginary part of a feature or feature result on the other hand.The feature extractor further comprises a raw feature compressor to perform a compression of a dynamic range for at least two raw feature components, in order to obtain at least two compressed raw feature components for each raw feature result. Thus, in one embodiment, the raw feature compressor not only compresses the magnitude of a raw feature result, but also compresses the phase of the raw feature result. Alternatively, the raw feature compressor not only compresses the real part of a raw feature result, but also compresses, in addition to the real part, the imaginary part of the raw feature result.

[0020] Both at least two raw feature components for each raw feature result are provided to the signal processor for processing the feature set that presents two compressed components, in order to obtain a processed information signal, where the information signal can be, for example, an audio signal, an image signal or a radar signal. Petition 870250093764, dated 10 / 14 / 2025, page 86 / 154 9 / 66

[0021] According to this second aspect, a preferred form of compression is power-law compression, which is used to compress / limit the dynamic range of each separate component of the input feature values, and the goal is to improve the efficiency and learning performance and generalization capacity of a deep neural network. Therefore, according to the second aspect, the preferred power-law compression is applied to the complete input feature representation, i.e., all components among a plurality of components, rather than just a part of the input, such as the magnitude component only, i.e., only a subset of the plurality of components.

[0022] This overall dynamic compression ensures that each representation of the feature set can be processed, that is, a real / imaginary representation can be processed as needed, or a magnitude / phase representation can be processed as needed. In this way, both representations can be processed, and it has been shown that the neural network processor's performance is better with respect to complexity and quality when compared to a situation in which only the magnitude is compressed, but not the phase.

[0023] According to the third aspect of the invention, multi-stage processing is performed in the neural network processor. An apparatus for processing an information signal, according to the third aspect of the present invention, comprises a feature extractor for extracting a set of features from the information signal, wherein each feature of the feature set comprises at least two feature components, wherein a first feature component among the at least two feature components is more important than a second feature component among the at least two feature components. Additionally, the Petition 870250093764, dated 10 / 14 / 2025, page 87 / 154 The 10 / 66 feature set comprises a first subset with the primary feature components and a second subset with the secondary, less important feature components.

[0024] The apparatus further comprises a neural network processor comprising a first neural network to receive, as an input, the first subset and to send a first processed subset. The neural network processor further comprises a combiner to combine the first subset and the second processed subset in order to obtain a combined subset. The combined subset is preferably fed into a second neural network to receive, as an input, this combined subset and to send a combined processed output. The combined processed output represents the processed information signal or the apparatus is configured to calculate the processed information signal using the combined processed output.In particular, the complexity of the first neural network that processes the first set of most important features is greater than the complexity of the second neural network that processes the second subset of less important features or the set of feature components.

[0025] Thus, according to the third aspect, which refers to a variety of applications related to deep neural network-based signal processing, the input resources are decomposed or divided into multiple components. Additionally, not all components need to be equally important in the context of the target task. Therefore, a DNN-based processing sequence, such as a neural network processor, is decomposed into multi-stage processing, where each stage handles the processing of components with decreasing importance in the context of the task at hand. In particular, each Petition 870250093764, dated 10 / 14 / 2025, page 88 / 154 The subsequent 11 / 66 stage deals with new input information, but less relevant or less important, so the neural networks or DNN modules in such a design may also have decreasing computational complexity from stage to stage.

[0026] Preferably, all aspects can be combined with each other, but all aspects can also be implemented separately from each other, or only two of three aspects can be combined, depending on the case.

[0027] In a combination of the first aspect and the second aspect, power-law compression is applied to the feature set that is registered in the feature segmenter. Thus, the feature set, which is further processed by the feature segmenter and the subsequent elements of the apparatus for processing an audio signal, according to the first aspect, is a feature set possessing feature components that are compressed with respect to their dynamic range, not just with a single component, but with two or more, i.e., multiple components, such as the magnitude and phase of the real and imaginary parts.

[0028] In such a situation, a feature decompression occurs before or after the feature combiner processing in the processing direction so that, in the end, the output post-processor post-processes the resulting decompressed feature set in order to obtain the processed information signal.

[0029] In such a combination of the first aspect and the second aspect, the third aspect can also be implemented in the first neural network or in the second neural network that processes the first and second feature subsets.

[0030] In this way, a combination of all three aspects is achieved, in which the resources used by the feature segmenter Petition 870250093764, dated 10 / 14 / 2025, page 89 / 154 12 / 66 are compressed, and at least one of the first neural network or the second neural network operates in a low-complexity, multi-stage mode, in which neural networks with decreasing complexity operate to process features with decreasing importance from stage to stage.

[0031] However, the first aspect of the invention can also be applied using the multi-stage processing of the third aspect, but without using power-law compression for both components, according to the second aspect.

[0032] In other embodiments, the second aspect of the invention may be carried out together with the third aspect or independently, but without the first aspect relating to overlapping feature segmentation and the corresponding feature combination with a low-complexity neural network. In such a procedure, the feature set may nevertheless be split along the channel direction, but without segmentation into overlapping segments. However, splitting along the channel direction may also occur with overlapping segments.

[0033] Additionally, the third aspect regarding low-complexity, multi-stage neural network processing can also be performed without feature segmentation or splitting along the channel direction, and even without power-law compression, depending on the case. However, combining the second aspect with the first aspect is preferred in the case of processing applications with moderate feature set sizes. Such moderate sizes occur, for example, in audio applications with relatively low audio sampling rates, resulting in relatively small feature set sizes. This is Petition 870250093764, dated 10 / 14 / 2025, pp. 90 / 154 13 / 66 illustratively shows the case where feature extraction occurs using time and frequency decomposition based on a bank of real-valued or, preferably, complex-valued filters, such as QMF and FFT-based filters, which generates a time and frequency representation of the information signal, such as an audio signal or radar signal. In image signals, feature extraction can be a spatial filter bank or spatial Fourier transform that generates, for an image, a one- or two-dimensional representation of cosine-based functions represented by an amplitude or phase, or a real and an imaginary part.

[0034] Preferred embodiments of the present invention are subsequently described with respect to the accompanying drawings, in which:

[0035] Figure 1a is a device or a method for processing an information signal, according to the first aspect;

[0036] Figure 1b is a device or a method for processing an information signal, according to the second aspect;

[0037] Figure 1c is a device or a method for processing an information signal, according to the third aspect;

[0038] Figure 2 is a schematic view of an audio processing sequence based on a deep neural network (DNN);

[0039] Figure 3a is an implementation of a feature extractor and a feature segmenter, according to the first aspect;

[0040] Figure 3b is an implementation of the resource combiner and output post-processor, according to the first aspect;

[0041] Figure 4 is a preferred implementation of the feature segmenter and sub-band splitting per channel performed subsequently, according to the preferred embodiment of the first Petition 870250093764, dated 10 / 14 / 2025, pp. 91 / 154 14 / 66 aspect;

[0042] Figure 5 is a preferred implementation of the neural network processor, according to the first aspect;

[0043] Figure 6 is an illustration of the segmentation of the feature set into a first feature set and a second feature subset;

[0044] Figure 7 is an illustration of the second subset of features before and after splitting into multiple segments arranged along a channel dimension;

[0045] Figure 8 is an implementation of the resource combiner comprising a stacker to perform a per-channel stacking and a subsequently connected frequency sub-band mixing;

[0046] Figure 9 is an illustration of the result of the second neural network in Figure 1a before and after stacking the segments;

[0047] Figure 10 is an illustration of the first subset of features and the second subset of features after the sub-band stacking by channel to the left in Figure 10, and the combination of the first result and the second result at the input of the third neural network to the right of Figure 10;

[0048] Figure 11 is a preferred implementation of the neural network processor, according to the third aspect;

[0049] Figure 12 is another preferred implementation of the neural network processor, according to the third aspect;

[0050] Figure 13 is a preferred implementation of the combinator for figures 1c, 11, or 12;

[0051] Figure 14a is a preferred implementation of the raw resource compressor of Figure 1b, according to the second aspect;

[0052] Figure 14b is a preferred implementation of a part of the signal processor in Figure 1b, according to the second aspect; Petition 870250093764, dated 10 / 14 / 2025, p. 92 / 154 15 / 66

[0053] Figure 15a is a preferred implementation of the present invention combining the first aspect and the second aspect;

[0054] Figure 15b is a table illustrating the complexity of the modality in Figure 15a and the modality in Figure 16;

[0055] Figure 16 is a preferred implementation of the invention, according to the second aspect, or the second aspect and the third aspect;

[0056] Figure 17 is another preferred implementation of the present invention, according to the second aspect, or the second aspect and the third aspect; and

[0057] Figure 18 is a table illustrating the complexity in MFlops of the modality shown in Figure 16 and two implementations of the modality shown in Figure 17.

[0058] Figure 1a illustrates an apparatus or method for processing an audio signal, according to the first aspect. The apparatus comprises a feature extractor 100 for extracting a set of features from the information signal, such as an audio signal, the set of features having a first dimension. The set of features having the first dimension is registered in a feature segmenter 200 to segment the set of features into a first subset of features and a second subset of features, wherein the first subset of features has a second dimension and the second subset of features has a third dimension, wherein the second and third dimensions are less than the first dimension. Obviously, a larger number of subsets, such as three or four, can also be used.Additionally, the first subset of resources and the second subset of resources have an overlapping range, such that one or more of the resources in the resource set are included in both the first subset of resources and the second subset of resources. Petition 870250093764, dated 10 / 14 / 2025, pp. 93 / 154 16 / 66 features. The first subset and the second subset of features are registered in a neural network processor 300. The neural network processor 300 is configured to process the first subset of features using a first neural network 310 in order to obtain a first result, and to process the second subset of features using a second neural network 320 in order to obtain the second result.

[0059] The first and second networks may be two different or independent networks, or they may be part of the same overall neural network, wherein parts may be separated from each other or both may use one or more parts of the overall neural network, provided that one or more parts of the overall neural network can be identified and used only by the first neural network 310 or by the second neural network 320.

[0060] Both the first and second results are recorded in the feature combiner 400 to combine the first and second results using a third neural network 420 illustrated in Figure 8. The third neural network has a complexity lower than the complexity of the first or second neural network. The result of processing by the feature combiner 400 is a resulting feature set that has a result dimension greater than the second or third dimension and preferably equal to the dimension of the feature set as obtained by the feature extractor 100. The resulting feature set is recorded in an output post-processor 500 to post-process the resulting feature set in order to obtain a processed audio signal.

[0061] In a preferred implementation, the feature extractor 100 comprises a time and frequency decomposer to generate, from the audio signal or radar signal, in a time-domain representation, a decomposed signal being a representation of Petition 870250093764, dated 10 / 14 / 2025, pp. 94 / 154 17 / 66 time and frequency comprising a sequence of time frames, where each time frame has a number of frequency compartments. Additionally, the feature extractor 100 comprises a feature set constructor to build the feature set from the time and frequency representation. Figure 6 illustrates, on the left, an extraction of a time and frequency representation with only six time frames and an illustrative number of only 12 frequency compartments, although many additional frequency compartments can, in fact, be used in practice, as will be discussed later. The example in Figure 6 already illustrates the situation subsequent to the operation of the feature of said constructor, since the sequence of six time frames has already been extracted from the larger time and frequency representation to obtain a feature set.Through the operation of the feature segmenter, the first subset of features illustrated in the lower right of Figure 6 and the second subset of features illustrated in the upper right of Figure 6 are obtained. Figure 6 illustrates that the overlapping range comprises two frequency compartments between 8 kHz and 12 kHz, although it is emphasized that this is only an example. In fact, for the purposes of a 48 kHz sampling rate resulting in frequency features up to 24 kHz, a time frame will comprise, for example, 769 frequency features.

[0062] In order to undo the feature extraction, the frequency and time composer is provided as a functionality of the output postprocessor 500 of Figure 1a which composes, from a set of input features in the time and frequency representation, the processed information signal being in the time domain representation. In one embodiment, the set of input features in the resulting feature set or, alternatively, the Petition 870250093764, dated 10 / 14 / 2025, pp. 95 / 154 18 / 66 The set of input features registered in the frequency and time composer, or frequency and time transformer, is derived from the resulting feature set and the feature set, that is, the features being registered in feature extractor 100 of Figure 1a.

[0063] When, for example, the present invention is applied to calculate the mask or time and frequency mask for noise reduction or speech enhancement purposes, the set of input features recorded in the time and frequency composer is not the mask itself, but the result, when the mask itself is applied to the original time and frequency representation, was generated by the feature extractor 100. Time and frequency masks can be applied as spectral gains to the input signal recorded in the feature extractor 100 in order to achieve the result of noise reduction or speech enhancement. Alternatively, the time and frequency mask can also be used to estimate, for example, the energy spectral density matrices of a noise or speech signal, and then one or more informed spatial filters can be calculated and then applied to the input signal.

[0064] When, however, processing is used, for example, for bandwidth extension purposes, then the result of the neural network processing, i.e., the resulting feature set as obtained by the feature combiner, may be the extended bandwidth audio signal in the bandwidth extension band or the total audio signal, i.e., the audio signal in the baseband and the extended band, but still existing in a time and frequency representation, and the functionality of the output post-processor 500 will simply be the frequency and time conversion of the extended bandwidth audio signal. When, however, neural network processing is applied to derive the extension parameters of Petition 870250093764, dated 10 / 14 / 2025, pp. 96 / 154 19 / 66 bandwidth from the input signal, the parameters must be applied to the input signal to derive, at the end of the extended bandwidth output signal.

[0065] Other applications, besides noise reduction or speech enhancement or bandwidth extension, may be source separation applications, where the input signal, for example, is an input signal that has its information, for example, from several different speakers arranged in different positions, and the input signal comprises, for example, several microphone signals. In this situation, the resulting set of features may, once again, consist of frequency masks, which may be binary, so that the value in the given time frequency compartment is one in the given compartment representing the time and frequency mask of the dominant source, and zero for the time and frequency masks representing a probability of the corresponding source being dominant in that compartment.As mentioned, time and frequency masks can be applied as spectral gains to one of the microphone signals in order to achieve source separation. However, the time and frequency mask can also be used as a means to estimate, for example, energy spectral density (PSD) matrices of the different source signals, and then the informed spatial filters are computed for the purposes of this blind source separation. Thus, depending on the determined implementation, the resulting feature set can be applied to the information signal and then the result is converted to the time domain, or the resulting feature set is already the processed signal in the frequency domain and the output post-processor only performs the processing to convert the frequency domain representation to the time domain representation. Petition 870250093764, dated 10 / 14 / 2025, pp. 97 / 154 20 / 66

[0066] Thus, for mask processing purposes, the output post-processor 500 comprises a mask processor for applying the time and frequency mask as a spectral gain mask to the feature set in the time and frequency representation in order to obtain the input feature set, or for computing a time and frequency mask processing filter and for applying the processing filter to the audio signal or to the feature set in order to obtain the input feature set.

[0067] In a further implementation, the first neural network is more complex than the second neural network, and the first feature subset comprises audio signal information from a lower frequency range of the audio signal or radar signal. Additionally, the second feature subset comprises audio or radar signal information from a higher frequency range of the audio or radar signal. Thus, the part of the audio or radar signal that is typically more important for perception or processing intent, respectively, i.e., the lower frequency range, is processed with a more complex neural network, and the part that is less important for speech perception, for example, of the audio signal, i.e., the higher frequency range of the audio signal, is processed using the second, less complex neural network 320 of Figure 1a.For radar signals, this situation can be different, as certain objects to be identified or located map to specific frequency ranges in the radar signals. For example, objects with certain velocities will exhibit Doppler frequencies in specific frequency bands that can be processed with greater complexity compared to other frequency bands that map to velocities that are not the primary focus of a surveillance task.

[0068] In a preferred embodiment, the resource segmenter Petition 870250093764, dated 10 / 14 / 2025, pp. 98 / 154 21 / 66 200 is already configured to generate the first subset of features and the second subset of features, such that the second dimension of the first subset of features is smaller than the third dimension of the second subset of features. In this way, a first subset of smaller dimension is processed with the first, more complex neural network 310, so that a large amount of processing energy is allocated to the specific part of the information signal in order to obtain the best possible refinement of the specific part, such as the lower frequency part, under the situation of limited resources.

[0069] The feature segmentation illustrated in Figure 6 can be further improved by splitting the second feature subset into multiple second segments. Additionally, these multiple second segments, obtained by means of, for example, the three split windows 1, 2, and 3 on the left side of Figure 7, are arranged along the channel dimension, as illustrated on the right side of Figure 7, so that the number of channels in the input set recorded in the second neural network 320 is improved. However, at the same time, when the number of channels is improved, the frequency or dimension of feature input is reduced.

[0070] In an additional embodiment, not only is the second subset, which, for example, covers the higher frequency part of the audio signal, divided, but also the first subset of features is divided into first multiple segments, and the first multiple segments are arranged along a channel direction to also optimize the number of channels of the input set to be recorded in the first neural network. As illustrated, for example, in the example shown in Figure 10, the first subset was divided into two segments, and the second subset was divided into three segments. In this way, the number of second multiple segments Petition 870250093764, dated 10 / 14 / 2025, pp. 99 / 154 22 / 66 equals three, and the number of the first multiple segments equals two. In other embodiments, and depending on the situation, it is also useful to perform only the frequency sub-band division per channel of the second subset of resources, i.e., the higher frequency band, as illustrated in Figure 7, and not perform such division of the first subset of resources, i.e., for example, for the lower frequency band of the audio signal. In such a situation, the lower part of Figure 10 will simply consist of a single green color. Preferably, and as illustrated in Figure 7, the division windows for dividing the subset of resources into the first or second corresponding multiple segments are overlapped, and particularly overlap by 50%, although a smaller overlap width can also be used.The same applies to the overlap applied by means of the frequency 200 sub-band segmentation illustrated in Figure 6, where an overlap band represents only one-third of the size of the entire subset.

[0071] In particular, Figure 6 illustrates the preferred implementation, wherein the first set of features ranges from about 0 to about 12 kHz, that is, it covers about 12 kHz of the frequency compartments, while the second subset of features, that is, the subset of features covering the higher frequency ranges from 8 kHz to 24 kHz, covers 16 kHz of the frequency compartments.

[0072] The output-side processing corresponding to the frequency sub-band division by channel 301 or 302, as illustrated in Figure 4, is the frequency sub-band stacking by channel 401, 402 illustrated in Figure 8. Additionally, when both frequency sub-band division by channel and frequency sub-band segmentation 200 are applied, then both features are undone on the output side and, for this reason, the Petition 870250093764, dated 10 / 14 / 2025, pp. 100 / 154 23 / 66 invention, according to the first aspect, uses a third neural network that has a lower complexity than the complexity of the first neural network or the second complexity of the second neural network and, preferably, the third neural network is even less complex than both the first neural network 310 of figure 5 and the second neural network 320 of figure 5.

[0073] According to the first aspect, this mixing is performed simply by stacking the first result and the second result and, if necessary, simply by stacking the individual channels of the results, so that a stacked feature set is obtained. Due to the overlap range of frequency sub-band segmentation and due to the preferred overlap range of frequency sub-band division by channel, the dimension at the input to the third neural network will be significantly larger than the dimension of the feature set, i.e., the first dimension at the output of the 100 feature extractor.Nevertheless, due to the fact that the third neural network only needs to undo the dimension increase and, with respect to the feature set, only needs to significantly process the feature set for an overlapping frequency range, while typically the features in a non-overlapping frequency range will not need to be heavily processed by the third neural network, this network can exhibit low complexity even in view of an increased input feature dimension. Therefore, it is preferable that not only is the complexity of the third neural network lower than the complexity of the first and second neural networks, but also that the number of layers of the third neural network is lower than the number of layers of the first and second neural networks, and particularly, it is preferable that the number of layers of the third neural network be at most 1 / 3 or more, preferably only 1 / 4. Petition 870250093764, dated 10 / 14 / 2025, pp. 101 / 154 24 / 66 or even more preferably only 1 / 10 of the number of layers of the first neural network or the second neural network.

[0074] The preferred idea in this respect is to introduce a resource redirection block (CR1 in figure 3a) in P1, after the resource computation step, which results in the use of efficient hyperparameter configurations of the DNN model block to reduce computation complexity.

[0075] Following the input processing by the DNN processing block (D1), the feature reorientation step is reversed in order to restore the DNN output to the original dimensions of the input features (CR2 in figure 3b) and, finally, the feature extraction step is reversed to obtain the output audio waveform.

[0076] As described earlier and illustrated in Figure 2, the preprocessing block P1 consists of a feature extraction / computation step, where the audio input signal is transformed into a representation that can, for example, be a 2D or 3D real-valued tensor. For example, when the applied feature extraction step is STFT, then the feature representation is a 2D complex-valued tensor that can be represented as a 3D real-valued tensor of size N x K x 2, where N corresponds to the number of frames, K corresponds to the number of frequency compartments, and 2 corresponds, for example, to the real and imaginary components or to the magnitude and phase components. In the case of the applied feature extraction step being a learned filter / feature bank, the feature representation is generally a 2D real-valued tensor of size N x K, where, in this case, K corresponds to the size of the feature dimension, while N would correspond to the number of time frames.

[0077] To learn from the local structure of the representation Petition 870250093764, dated 10 / 14 / 2025, pp. 102 / 154 25 / 66 feature layers, convoluted layers with small convolutional cores are typically applied to the initial stages of a DNN model for audio processing. Due to the use of multiple such layers, the computational complexity of the model is primarily affected by the spatial dimensions of the input features, N and K. Alternatively, other layer types, such as fully connected or recurrent layers, can also be used in the initial stages of the DNN model, and in this case, the main factors affecting complexity would also be N and K.

[0078] The number of time frames N is determined by the window / frame length considered for the transformation / filter bank / feature extraction applied to the signal in addition to the length of the recorded audio signal. In real-time audio applications, which is the focus here, very small audio segments, typically less than 50 ms, are processed at a time, resulting in N being small for each processing case. Therefore, the focus in this invention is basically on the frequency / feature dimension of the input features.

[0079] The size of the feature / frequency dimension K is determined by the desired frequency / feature resolution in addition to the audio signal sampling rate. For speech or audio processing tasks, the feature / frequency dimension can, for example, be selected as a power of 2 and greater than 128. For higher sampling rates, such as 32 or 48 kHz, the size of the feature dimension is typically very high, for example, 512, 1024, etc. To reduce the computational complexity of the main DNN model when the size of the feature / frequency dimension is very high, the invention presents a combination of two feature reorientation techniques for complexity reduction, as illustrated in Figure 4. The techniques are described below considering the use of Petition 870250093764, dated 10 / 14 / 2025, pp. 103 / 154 26 / 66 a deterministic transformation / filter bank, such as STFT, as the feature extraction step. The following descriptions explain the methods considering the feature dimension as being the frequency dimension specifically; however, the methods can also be applied in the case of using any arbitrary feature representation. Sub-band segmentation

[0080] The first technique corresponds to a frequency sub-band segmentation method in which the input feature is divided along the frequency dimension K into multiple overlapping segments, [si, S2,...,Sm). In some academic works, a basic non-overlapping division idea has been presented to improve the performance of speech enhancement methods based on DNN specifically [1,2]. In this report, the segmentation approach is presented as a method to reduce complexity and extends to be an overlapping division.

[0081] Following the division of the original input resource into segments, each resource segment is prepared as an input to an individual DNN model for processing each of the segments separately. Therefore, in these cases, instead of a single DNN model, block D1 in Figure 1 now consists of multiple DNN models (as illustrated in Figure 5) that operate separately on different frequency sub-band segments created by the sub-band segmentation technique.

[0082] An example of the segmentation technique and the creation of corresponding input segments for individual DNNs is illustrated in Figure 6. In the example provided, an audio signal with a sampling frequency of 48 kHz was used. The computation of STFT features provides a feature representation where the frequency dimension represents the signal content up to the frequency. Petition 870250093764, dated 10 / 14 / 2025, pp. 104 / 154 27 / 66 Nyquist frequency of 24 kHz. With the division method presented, the frequency dimension in this case is segmented into two parts: a feature representation for the lower frequency range of 0 to 12 kHz, and another for the higher frequency range of 8 to 24 kHz. Therefore, in this example, these two segments form the input of two different DNN models in block D1. * Sub-band segmentation can be chosen arbitrarily. Depending on the desired frequency dimension size for the individual DNN inputs and the number of individual DNNs to be employed, the number of M segments and the frequency dimension size for each segment can be chosen. * The frequency dimension size does not need to be the same for different segments. Depending on the audio processing task, the size of individual segments can be determined. For example, for speech processing tasks, the size of the first segment can be determined based on the typical frequency range of speech activity when STFT is used as the feature representation. The size of the remaining segments can be determined based on the number of DNNs that can be employed. Alternatively, band segmentation can be performed using equivalent rectangular bandwidth (ERB), the Mel scale, or similar perceptual scales depending on the task for which the method is being developed. * The overlap between segments avoids processing artifacts in the boundary regions of the segmented feature representations when the processed outputs are blended to obtain the final audio output. Segmenting the input along the frequency dimension facilitates the use of individual DNNs with less computational complexity compared to a single large DNN model. Petition 870250093764, dated 10 / 14 / 2025, pp. 105 / 154 28 / 66 A singular model, due to the large input space, has a very high number of computations per layer, especially in the initial layers, resulting in a high overall computational complexity of the system. By segmenting the input along the frequency dimension, the computational complexity in the initial layers can be controlled to be low enough to keep the overall system complexity low. The flexibility in determining the segmentation allows the designation of an overall system in which the combined complexity of individual DNNs is less than that of a single DNN model. * On hardware platforms that allow parallel processing, processing with individual DNNs can be performed in parallel to further reduce the overall system computation time. Sub-band division by channel

[0083] For this method, it is considered that the initial layers of the individual DNN models (figure 5) consist of convolution layers that use small convolution kernels to learn from the local structure of the audio input feature representation. This is generally true for most audio processing DNN models, where exploration / learning of the audio input feature representation from the local structure is important for improved performance across a wide variety of tasks.

[0084] In this method, the individual input segments obtained after the Sub-Band Segmentation block are further divided into multiple overlapping segments of equivalent spatial dimensions and arranged along the channel dimension. The idea of ​​channel splitting has been previously proposed in [3, 4] as a more efficient feature extraction step compared to applying Petition 870250093764, dated 10 / 14 / 2025, pp. 106 / 154 29 / 66 convolution through the entire input space. In the case of this invention, the idea is to use it in conjunction with the sub-band segmentation method as the second stage of complexity reduction. As mentioned earlier, sub-band segmentation allows the maintenance of low computational complexity in the previous convolution layers due to the reduced input space (NxK -> Nxki, NxK2,..., where Ki, K2,..., <K). A divisão por canal estende, ainda, esse benefício através da redução adicional das dimensões espaciais (NxKi, NxK2,...) do espaço de entrada de cada uma das entradas individuais.

[0085] As an example, consider the segmented blocks on the right side of Figure 6. Then, the segmented block, which corresponds to the higher frequency features, can be further divided into three uniform overlapping segments, and each divided segment is then laid out along the channel dimension in order to obtain the input feature representation for the individual DNN.

[0086] -The overlap factor of the split can be chosen arbitrarily, under the constraint that the spatial dimensions (number of time frames and size of the frequency sub-band dimension) of the individual splits need to be equal. This is necessary so that they can be arranged along the channel dimension. In a typical convolution layer operation, small filters are applied to each channel separately, followed by a weighted combination of the filtered values ​​across the channel dimension in order to generate the output. Due to this fact, channel splitting assists in combining the cross-frequency sub-band information in the initial layer of the DNN, while presenting smaller spatial dimensions of the input for small filters to traverse and reducing computational complexity in these initial layers.

[0087] -The overlap factor can be determined based on Petition 870250093764, dated 10 / 14 / 2025, pp. 107 / 154 30 / 66 desired reduction of the spatial dimensions of the input and the number of channels. For example, in Figure 7, to reduce the number of channels in the DNN input, the input can be divided into three segments of equal spatial dimensions in order to reduce the operational complexity in the first convolution layer. Mixing Methods

[0088] Following the processing of the inputs by the individual DNNs, at the output of each DNN in D1, an improved / modified version of the input feature representation is obtained. The next task would be to recombine / mix the individual outputs in order to obtain the output features in their original dimensions, so that they can be provided as an input to the feature inversion step (inverse STFT in this case) to obtain the output audio signal.

[0089] In previous works employing sub-band segmentation [1,2] or sub-band splitting by channel [3,4,5], the mixing method is generally an approach in which different sub-band outputs are concatenated along the geometric frequency axis in order to obtain the final output. This is generally possible since the methods do not consider overlapping segmentation strategies. In one work [5], an approach was preferred, in which the concatenated output traversed a separate speech enhancement model for a two-stage enhancement approach. In this case, the second enhancement model presents the same complexity as the first.

[0090] In this invention, a simple concatenation along the frequency dimension is used, followed by a small neural network to learn the ideal combination / mixture of cross-frequency information, as well as to compensate for overlap between segments.

[0091] The two mixing blocks within the general mixing method Petition 870250093764, dated 10 / 14 / 2025, pp. 108 / 154 31 / 66 are illustrated in figure 8. Sub-Band Mixing per Channel * The outputs obtained from the individual DNNs are not restricted to any specific dimensions. The spatial and channel dimensions of these outputs can be arbitrary and different for each DNN. Each of the output channels is stacked along the feature / frequency dimensions to obtain the input for the next mixing block. It is important to note that the overlap considered during the split is not replicated in the stacking step. * Considering the split segments in Figure 7, the mixing stage is as illustrated in Figure 9. It is important to note that the frequency dimension of the mixed output is always greater than or equal to the frequency dimension of the original input (Figure 7), since any overlap between the split segments is not compensated for during mixing. Sub-band frequency mixing

[0092] For the frequency sub-band mixing stage, the individual outputs, after the previous mixing stage, are stacked in a similar way to that described for the channel sub-band mixing stage.

[0093] Figure 10 illustrates this mixing step of the two sub-band outputs that were obtained.

[0094] Following this first step, a small neural network is used as the final mixing step in order to obtain the representation of the output feature in a desired dimension, such as the original dimension of the input to the system, or in a desired dimension different from the original dimension of the input to the system, such as a dimension smaller than the original dimension. The neural network employed is not constrained in such a way that Petition 870250093764, dated 10 / 14 / 2025, pp. 109 / 154 32 / 66 presenting any specific architectural blocks, it can vary between being a multilayer perceptron (MLP) network with few fully connected layers and being a neural network that is a combination of convolutional, recurrent, and fully connected layers. The only constraint of the neural network model is that the last layer of the neural network must rearrange the frequency dimension of the output so that it is equal to the frequency dimension of the original input to the audio processing system. This is necessary to reverse the transformation step. Alternatively, the last layer rearranges the frequency dimension of the output to be equal to a desired dimension that is different, such as smaller, than the frequency dimension of the original input to the audio processing system. This may be the case when a classifier is present. Modalities

[0095] 1. In the first embodiment, all processing modules described as part of this invention can, for example, be applied to the design of a low-complexity DNN based on the audio processing system.

[0096] 2. In the second embodiment, all processing modules, except for sub-band segmentation, can be used to design a low-complexity DNN based on the audio processing system. Sub-band segmentation can be ignored in cases where the feature dimension is not very large, for example, when the feature dimension is less than 512.

[0097] 3. Additional embodiments may include all other possible combinations of different processing modules described herein.

[0098] Subsequently, a preferred modality of the second Petition 870250093764, dated 10 / 14 / 2025, pp. 110 / 154 33 / 66 An aspect related to the generation of compressed components, for example, using power-law compression, is discussed with regard to Figure 1b and Figures 14a, 14b.

[0099] A device for processing an information signal, which may be an audio signal, but may also be an image signal, or a radar signal, for example, comprises the feature extractor 100 basically as illustrated previously with respect to figures 1a to 10. In particular, the feature extractor 100 is configured to extract a set of features from the information signal, and the feature extractor comprises a raw feature calculator 120 to calculate the raw feature results, wherein such raw feature result has at least two raw feature components as illustrated in figure 1b below block 110. One raw feature component of a raw feature result may be a magnitude and the other raw feature component of the raw feature result may be the phase of a complex amplitude of a cosine-type function, for example.Alternatively, the first component could be the actual part and the second component could be the imaginary part of the corresponding raw resource result.

[00100] The two components, that is, the magnitude and the phase, or the real part and the imaginary part, for example, are registered in a raw resource compressor 120 to perform a compression of at least two raw resource components in order to obtain at least two compressed raw resource components for each raw resource result as illustrated by the two arrows below block 120 in figure 1b. In particular, the resource set sent by the resource extractor therefore comprises the compressed raw resource components for each raw resource result.

[00101] These raw resource components in the form Petition 870250093764, dated 10 / 14 / 2025, pp. 111 / 154 Compressed 34 / 66 data are then registered in a signal processor 600 for processing the feature set comprising the compressed raw feature component for each raw feature result in order to obtain a processed information signal. The signal processor 600 may comprise one or more neural networks, as discussed previously with respect to the first aspect, as in Figure 1a or Figure 5, or it may comprise a single neural network 20, as illustrated in Figure 2, or it may comprise cascaded neural networks, as illustrated with respect to the third aspect illustrated in Figure 1c or Figures 11 to 13. The signal processor may additionally comprise a feature segmenter, a feature combiner and / or a post-processor as was, for example, illustrated with respect to Figure 1a and the related description.

[00102] On the other hand, the resource extractor 100 in figure 1a and the resource extractor, as illustrated in figure 3a, can also be implemented as illustrated in figure 1b, that is, as discussed with regard to the second aspect.

[00103] The raw resource calculator 110 is preferably configured to calculate the complex value as the raw resource result, the raw resource result being a real part or an imaginary part, or the magnitude and phase, provided that at least two raw resource components, and the raw resource compressor is configured to perform a compression of the real part and the imaginary part or the magnitude and phase to obtain the compressed raw resource components.

[00104] The raw resource compressor 120 is additionally configured to apply 121 a first compression function to a first raw resource component and apply 122 a second compression function to a second raw resource component, wherein the second compression function may be different from the first. Petition 870250093764, dated 10 / 14 / 2025, pp. 112 / 154 35 / 66 compression function. Preferably, the raw resource compressor is configured to apply a power-law compression using a power value of 1 / α, where α is a fixed value or is a specific value learned from a specific application and, in particular, α consists of a real or integer number greater than 1. Alternatively, the power value is α. Then, the power value α is a fixed value or is a specific value learned from a specific application and, in particular, α consists of a real or integer number less than 1.

[00105] Preferably, the α number is different for each raw feature component and is particularly higher for raw feature components that should exhibit a higher numerical range, such as for magnitude and phase. Preferably, the compression function is stronger with respect to its magnitude compression action compared to its phase compression, but when the real and imaginary parts are expected to exhibit similar value ranges, similar compression function intensities and therefore similar α numbers for parallel compressions can be applied.

[00106] Preferably, the raw resource calculator 110 is configured to calculate the raw resource components with absolute numbers and associated signs, and the raw resource compressor 112 is configured to apply the compression function corresponding to the absolute value and retain the sign of a corresponding raw resource component.

[00107] Preferably, the information signal is an audio signal, and the raw resource calculator 110 comprises a time and frequency decomposer to decompose the audio signal into a complex time and frequency representation. Alternatively, the information signal is an image signal, and the raw resource calculator 110 comprises a spatial transformer to decompose the Petition 870250093764, dated 10 / 14 / 2025, pp. 113 / 154 36 / 66 image signal in a complex spatial frequency representation.

[00108] Preferably, the signal processor 600 in Figure 1b comprises a neural network processor 300, as illustrated in Figure 1a or 5 or in Figure 2, with reference number 20 in order to obtain a processed feature set. Additionally, as illustrated in Figure 14b, a feature decompressor 350 to perform a decompression that coincides with the compression performed by the raw feature compressor 120 is provided to obtain a resulting feature set. Additionally, a post-processor 500 is configured to post-process the resulting feature set in order to obtain the processed information signal.

[00109] Particularly, depending on the implementation and, in particular, depending on whether the 200 feature segmentation and 400 feature combination of Figure 1a are applied, it is preferable that the 350 feature decompression occur after the 400 feature combination, i.e., the third neural network with a low complexity operating on the compressed features as obtained by the 300 neural network processor. However, in other embodiments, the feature combination can also occur in the uncompressed domain or in the real-imaginary part of the domain as described, so that the decomposed 350 feature is placed between the 300 neural network processor and the 400 feature combiner. To illustrate different alternatives, the first 350 feature decomposer is illustrated as being optional by means of the dotted line and the same applies to the 350 feature decompressor which can also be passed by the dotted line.

[00110] In another preferred embodiment, which is illustrated and described later with regard to Figure 16 and Figure 17, the functionality of raw resource compression 120 and law decompression of Petition 870250093764, dated 10 / 14 / 2025, pp. 114 / 154 37 / 66 subsequent power 350 are combined with processing using a single DNN 20, as illustrated in Figure 2. Additionally, in Figure 17, the power-law compression feature, illustrated according to the second aspect of Figure 1b, is also combined with the functionality of sub-band division per channel 301 and sub-band mixing or stacking per channel 401, as discussed previously with regard to Figure 4 and Figure 8, but without the functionality of the first aspect, i.e., sub-band segmentation 200 and frequency sub-band mixing in a small post-processing neural network 420.

[00111] Additionally, the functionality of relying on a single neural network instead of two parallel neural networks, as illustrated in Figure 5, Figure 1a, or Figure 15a, is preferred for audio signals with sampling rates below 24 kHz and where the feature set dimension in the frequency dimension is less than 500 and greater than 200, and the number of channels is equal to 2. In such a situation, it is preferable to use a single neural network, such as the 20d neural network in Figure 2, or to use the cascaded neural network, as discussed in relation to the third aspect of Figure 1c or in the context of Figures 11 to 13.

[00112] Additionally, not only for this aspect but also for the other aspects, it is preferable to use a number of segments for channel division between 2 and 4, so that the number of channels is between 4 and 8, and the overlap of segments for channel division is between 30 and 90 frequency compartments and, preferably, between 60 and 70 frequency compartments.

[00113] In this descriptive report, power-law compression is used to compress / limit the dynamic range of each separate component of the input feature values, and the goal is to improve the learning and generalization capability of the DNN, i.e., the Petition 870250093764, dated 10 / 14 / 2025, pp. 115 / 154 38 / 66 Power-law compression is applied to the complete representation of the input features rather than just a portion of the input, such as the magnitude component. In mathematical terms, the compressed signal with the proposed power-law compression can be expressed, for example, as: Ck = Ck,re + jCk,im = Ck,abse - jCk,ph

[00114] Where Ck,re is the compressed real part component, Ck,im is the compressed imaginary part component, Ck,abs is the compressed magnitude component and Ck,ph is the compressed phase component of the input signal k, respectively.

[00115] In general, different compression functions can be used for each component and input channel k. The compression of a specific component and channel k can be expressed as follows: CM.) - fk,(.>(Sk,(.))

[00116] Where f denotes an arbitrary compression function and the subscript (·) indicates different signal components, for example, the real component re or the imaginary component im. After processing the DNN, the power-law compression can be reversed to obtain the components of the first final representation of the output feature, that is, rfc,(·) = f—)(Dk,(·))

[00117] Where f-1 represents the inverse power law function (expansion function). The final output feature representation of the k-channel can then be obtained from the individual components, for example: / fc = / fc,re + j / k,im = Tfc,abse^k,ph

[00118] Where Tk is the element k of the vector t.

[00119] As an illustrative implementation of the compression steps above, one can consider the representation of real and imaginary signals. In this case, the compressed components can be Petition 870250093764, dated 10 / 14 / 2025, pp. 116 / 154 39 / 66 obtained as: Ck,re= (±)\Sk^e Oc.im = (±)\5k,imAm

[00120] Where (+ / -) denotes the original sign of the feature value Sk,(->. As can be observed, in this example, the power-law function fx,(.> corresponds to obtaining the absolute value of the specific signal component while retaining the signal information. Furthermore, power-law compression is applied to the complete signal representation, that is, to both the real and imaginary parts. After processing with the DNN, the power-law compression can be inverted accordingly in order to obtain the final output feature representation components, that is, Tk,(·) = (±)|Ok.(.)l1 / “'-(')

[00121] As another illustrative implementation, magnitude and phase representation can also be considered. In this case, compressed signal components can be obtained as: of the Gc.abs = (5k,abs)“k,abS Oc.ph = $k.ph · «k.ph

[00122] In this case, although similar to the previous example, the power-law compression function is applied to the complete signal representation; the function fk,(.) is different for the two components in this case. After processing with the DNN, the power-law compression is inverted accordingly, that is, Τ' _ \1 / afc,abs ^k.abs = (^k.abs) Tfc.ph = ^k.ph / ^fc.ph to obtain the final output resource representation.

[00123] In the proposed power law compression, * the power law function f (or, for example, the power law factor α) can be Petition 870250093764, dated 10 / 14 / 2025, pp. 117 / 154 40 / 66 * different for each component (e.g., re, im, abs, ph) and the k element of the input feature representation; * α can be fixed or learned from a specific signal processing task through training.

[00124] Subsequently, reference is made to Figure 1c in order to illustrate the third aspect of the invention relating to low-complexity, multi-stage DNN processing. An apparatus or method for processing an information signal is illustrated in Figure 1c, and the apparatus comprises a feature extractor 100 for extracting a set of features for the information signal. Each feature of the feature set comprises at least two feature components, wherein a first feature component of the at least two feature components is more important than a second feature component of the at least two feature components, and wherein the feature set comprises a first subset with the first feature components and a second subset with the second feature components.

[00125] Illustratively, the most important components are the magnitude of the spectral values ​​obtained by a time and frequency decomposition of the information signal, for example, an audio signal, an image signal, a radar signal, or any other information signal that can be subjected to feature extraction resulting in components that are more or less important for an individual feature.

[00126] The apparatus further comprises a neural network processor 300 having a first neural network 340 to receive, as an input, the first subset 322 and to send a first processed subset. Additionally, the neural network processor comprises a combiner 350 to combine the first processed subset and the second subset 331 in order to obtain Petition 870250093764, dated 10 / 14 / 2025, pp. 118 / 154 41 / 66 a combined subset. Additionally, the combined subset, now presenting, for each feature, the first and second processed components, is registered in a second neural network 360 which sends a processed combined output that is registered in an optional output stage 700. Specifically, the combined output represents the processed information signal or the device is configured to calculate, using the output stage 700, the processed information signal using the combined output.

[00127] Additionally, the complexity of the first neural network 340 is greater than the complexity of the second neural network 360. Thus, the most important components are recorded in the more complex neural network 340 and the less important components 331 are recorded, after combination, in the second neural network of lower complexity.

[00128] Additionally, with regard to image processing, the foreground of the image will represent the most important components and the background of the image will provide the less important components. Then, the feature extractor will process the features so that, for a given image range, for example, one more important component and one less important component are calculated.

[00129] In a preferred embodiment, and this is also illustrated in Figure 13, the 350 combiner concatenates the first processed subset and the second subset along the channel direction so that, for example, the first processed set, which represents a magnitude mask, and the second set of less important components being the phase component are arranged together along the channel direction and are recorded in the second neural network which computes, in contrast to the first neural network, and records with both channels while the first neural network Petition 870250093764, dated 10 / 14 / 2025, pp. 119 / 154 42 / 66 340 calculates an input with only a single channel.

[00130] When the DNNm 340 in Figure 12 performs a split and concatenation along the channel dimension with an overlap, it is necessary to verify that the overlap is also generated in the second subset fed to the combiner 350, so that a combination of the stacked results of the network 340 and the corresponding stacked results of the second subset can be performed in the container 350. Alternatively, the output of the network 350 is processed in such a way that the overlap resulting from the split along the channel direction is eliminated, so that the second subset fed to the combiner 350 coincides with the output of the network 350 in order to perform the combination illustrated in Figure 13 or Figure 12.

[00131] Generally, the number of channels processed by the downstream neural network is higher than the number of channels processed by the lower-order neural network, that is, the neural network that is upstream in the processing cascade. A generalized situation of the 300 neural network processor is illustrated in Figure 11, where three neural networks 340, 360, and 380 are shown, in which the 340 neural network is the most complex, the 360 ​​neural network has medium complexity, and the 380 neural network has the lowest complexity. At the same time, the number of channels processed by the first 340 neural network is lower than the number of channels processed by the second 360 neural network, and the number of channels processed by the last 380 neural network is the highest when the 350 and 370 combiners always perform a concatenation along the channel, as illustrated in Figure 13.

[00132] Figure 12 illustrates a modality with only two neural networks 340, 360, in which the output of the second neural network 360 is a representation with a real component and an imaginary component, Petition 870250093764, dated 10 / 14 / 2025, pp. 120 / 154 43 / 66 while the combined input in the 360 ​​neural network is the magnitude and phase arranged along the channel direction, as illustrated in Figure 13.

[00133] Preferably, cascaded processing with two different networks having different complexities is preferred for the implementations illustrated in Figures 16 and 17, wherein the sub-band segmentation 200 and the corresponding frequency sub-band mixing 410 illustrated in Figure 15a, which requires two parallel neural networks 310, 320, are not processed. However, an additional channel sub-band division 301 and a corresponding channel sub-band mixing or stacking 401 as illustrated, for example, in Figure 17, are preferably combined with sequential or cascaded processing, as illustrated in Figure 11 and Figure 12, so that the main DNN 330 of Figure 17 is implemented as illustrated in Figure 11 and Figure 12, i.e., by means of the processing illustrated with reference numbers 340 to 380 or 340 to 360.

[00134] In a variety of applications related to DNN-based signal processing, input features can potentially be decomposed / divided into multiple components. Additionally, not all components need to be equally important in the context of the target task. In such cases, the DNN-based processing sequence can be decomposed into a multi-stage process with each stage handling the processing of components with decreasing importance in the context of the task at hand. If each subsequent stage handles new, but less relevant / important input information, the DNN modules in such a design can also exhibit decreasing computational complexity.

[00135] A block diagram of such a processing sequence Petition 870250093764, dated 10 / 14 / 2025, pp. 121 / 154 44 / 66 is illustrated in Figure 11. The input resource in this case can be decomposed into N components, the number according to the importance / relevance of the task. As illustrated, the first DNN only considers the most important / relevant resource component as the input and produces an intermediate output that is then combined with the second most important / relevant resource to be provided as input for the next DNN, which is of lower complexity than the first. This processing chain is continued for all N components of the input resource, with the output of DNN N considered the final output.

[00136] For a more specific description, one can consider the application of noise reduction and the use of STFT as the feature extraction step. The idea can also be applied to other signal processing tasks where the extracted feature can be decomposed into two or more individual components, for example, in image processing for certain enhancement tasks it may be preferable to process the foreground and background of the image separately in two steps.

[00137] In the case of STFT, which is a complex transformation, the resource representation can be decomposed into magnitude and phase components, that is, X = Xm* e-Jxp

[00138] Where Xm and Xp are the magnitude and phase components of the complex STFT representation. For the specific case of noise reduction, it is generally more critical to optimize the magnitude component. This fact is used to further reduce the complexity of the overall system.

[00139] Recognizing the importance of the magnitude component, in the first stage of processing, illustrated in the left side of Figure 12, a DNN model (DNNm) is used to estimate Petition 870250093764, dated 10 / 14 / 2025, pp. 122 / 154 45 / 66 a magnitude mask that can be used to refine the magnitude component of the input feature. This model can be designed to be of much lower complexity than a model that operates on both complex components of the input feature together. The computed magnitude mask is then combined with the phase component of the input feature at this stage.

[00140] If the dimensions of the magnitude mask and the phase component are equal, the combination can be a simple concatenation of two components along the channel dimension, as illustrated in Figure 13.

[00141] Alternatively, the phase can be combined with the magnitude to obtain the real and imaginary components separately, for example, Yr = Mm* cos (Xp), Yi = Mm* sin (Xp), where YR and Yl are the real and imaginary components of the intermediate output, MM is the magnitude mask obtained from DNNM, and XP is the phase component of the original input feature.

[00142] Once computed, the real and imaginary components can also be concatenated along the channel dimension as illustrated in Figure 13 to form the input for the next stage.

[00143] In the next stage (according to Figure 12), the combined magnitude mask and phase component are processed using another DNN model (DNNMP) in order to obtain the real and imaginary components of the final enhanced output signal, which is then fed into the inverse transformation to obtain the enhanced audio signal. Since, in this second stage, the main task is to enhance the phase along with the enhanced magnitude mask, a DNN model with even lower complexity than DNNM can be used to further reduce the Petition 870250093764, dated 10 / 14 / 2025, pp. 123 / 154 46 / 66 computational complexity of the system as a whole.

[00144] Figure 16 illustrates an implementation of a processing sequence, according to a second aspect, which includes 120 power-law compression and 350 power-law decompression. The DNN block can be implemented as illustrated in any of the figures in this application and can also be implemented as a direct neural network without subband segmentation, subband division by channel, and corresponding mixing methods, such as subband mixing by channel and frequency subband mixing. Such a DNN with sequential processing that includes, for example, convolution layers, LSTs, and convolution transpose layers for mask estimation related to noise reduction in a 48 kHz sampling frequency rate task can be implemented as illustrated in Figure 16.Specifically, the frequency dimension will be 769 per time frame, and the number of C channels will be equal to 2, since the input resources are complex resources and comprise a real part and an imaginary part, or a magnitude and a phase.

[00145] It was observed that, depending on the batch number B, and the time number T, which can be selected as necessary, this DNN demonstrated a computational complexity of 499.5 MFlops (mega flops) with a parameter size of 2.67 M (million). A significant portion of the computational complexity for this DNN architecture was contributed by the convolution and convolution transposition layers, since in this case, the convolutional core needs to move through a large geometric frequency axis which, in the embodiment of Figure 16, has a value of F = 769. In such a situation, the present invention, according to the first aspect, is highly beneficial. Since the sub-band segmentation method segments the geometric axis feature of Petition 870250093764, dated 10 / 14 / 2025, pp. 124 / 154 47 / 66 frequency with two overlapping segments through reduction of the geometric frequency axis dimension, and the sub-band-by-channel division method further drastically reduces the geometric frequency axis dimension of the segmented features by stacking fixed overlapping segments in one channel dimension. The device, as illustrated in Figure 15a, is obtained with the corresponding channel and frequency numbers. It appears, from Figure 15a, that the frequency dimension of the first subset is reduced from 769 to 385 and that the frequency dimension of the second subset is reduced from 769 to 513. It should be emphasized that the dimension is smaller for the first sub-band when compared to the second sub-band, and that the sum of the frequency dimension, i.e., 385 + 513, equals 898 frequency compartments compared to 769 frequency compartments in block 120 of Figure 15a. This is due to the overlapping band.

[00146] Additionally, as illustrated in Figure 15a, the lower sub-band, a two-stage sub-band split, is performed, while for the upper sub-band, a three-stage sub-band split is performed. In the left branch of Figure 15a, the number of channels is increased from 2 to 4, indicating two segments, while in the right branch, the number of channels is increased from 2 to 6, indicating three segments.

[00147] The corresponding mixture in block 401 results in 257 x 2 = 514 frequency values. The right-hand side results in 257 x 3 = 771 frequency values.

[00148] After frequency sub-band mixing, the frequency dimension is equal to 1285 with a number of channels equal to 2, and this dimension is reduced by the small post-processing neural network 422, the input dimension of 769 with two channels, which is the same figure as in block 120. Petition 870250093764, dated 10 / 14 / 2025, pp. 125 / 154 48 / 66

[00149] Power law compression in block 350 does not change anything with respect to dimension, and therefore the input features that can be registered in the post-processor have a channel dimension of 769 with two channels for a real / imaginary or magnitude / phase representation.

[00150] In the aforementioned DNN branch system 1 and 2, for comparison purposes, the same DNN architecture illustrated in Figure 16 for the sequential system was used. However, the parallel system described in Figure 15a has a computational complexity of 399.5 MFlops, which is 100 MFlops less than the system described in Figure 16. This is mainly due to the reduced geometric frequency axis dimension of the input resources, which represents a narrowing in the sequential system, achieved by the Sub-Band Segmentation and Sub-Band Division per channel method.

[00151] In practice, two separate, less complex DNNs can be used for branch DNN 1 and 2 compared to the DNNs used in Figure 15a. This is because, after the segmentation methods, the DNN needs to process more localized overlapping features than the DNN employed in Figure 16, so a simpler DNN can also perform the task efficiently. Another reason is that, in practice, higher frequency compartments have low energy, so higher frequencies contribute little to the subjective and objective improvement of, for example, mask estimation for the noise reduction task compared to lower frequency compartments, which contain higher energy and thus have more significance. Therefore, the DNN employed in branch 2 can be even less complex than branch 1 of the DNN.

[00152] Thus, two different DNNs for the branching of Petition 870250093764, dated 10 / 14 / 2025, pp. 126 / 154 49 / 66 DNNs 1 and 2 are used. The computational complexity of the DNN 1 and 2 branching methods are 112.5 MFlops and 96.3 MFlops, respectively, with parameter sizes of 2.57M and 2.42M, respectively. After the blending methods, a small DNN with a computational complexity of 25 MFlops and 3.18M parameters is used.

[00153] In total, the system has a computational complexity of 233.83 MFlops with 8.17+M parameters. This illustrates a reduction of 265.67 MFlops, a reduction in complexity compared to the sequential system described in Figure 16, even after having 3 times the number of parameters.

[00154] The performance of this system is also subjectively and objectively much better than the system described in Figure 16. It is evident that the inventors mentioned above make this system computationally efficient without compromising performance.

[00155] Figure 15b illustrates a comparison of the complexity of the modality in Figure 16 and the modality in Figure 16a, illustrating a drastic reduction in complexity as measured in MFlops.

[00156] Figure 17 illustrates another preferred embodiment that can be used, for example, for an application with a reduced sampling rate. While in the description above of Figure 15a a use case of an audio processing system for a sampling rate of 48 kHz was described, Figure 17 illustrates a use case that can be applied to a sampling rate of 16 kHz, where subband division by channel 301, along with subband mixing by channel 401, is used to reduce computational complexity for this lower sampling rate. An illustration of a 16 kHz sampling rate is also shown in Figure 16, but the implementation in Figure 17 illustrates a further reduction in complexity as illustrated in the table in Figure 18. In particular, the frequency dimension of 257 is reduced by a 3-fold subband division in block 301, of Petition 870250093764, dated 10 / 14 / 2025, pp. 127 / 154 50 / 66 so that the number of channels is increased from 2 to 6, and the frequency dimension of each channel 129 is at the number 387, which is greater than 257 due to the overlap applied to perform the sub-band division per channel 301. The sub-band mixing per channel illustrated in block 401 reduces the number of channels by 3 and increases the frequency dimension to the maximum of 387 for the frequency input feature dimension. The small post-processing neural network 420 reduces the dimension from 387 to the original 257 as illustrated in block 120, and the number is not altered by the power-law decompression procedure in block 350.

[00157] In the system, the original input resources have F = 257 and C = 2. The Channel Sub-Band Split method is implemented with a bin split window of 129 F and an overlap of 64 bins. Therefore, the output of the Channel Sub-Band Split block has a format (B, T, 129, 6). This system has a computational complexity of 48.23 MFlops. A similar system architecture without this split-mix method will have a computational complexity of at least 150 MFlops.

[00158] Computational complexity can be further reduced by using only the smaller split window and overlap. In the least computationally complex model with this architecture (with minor changes to the RNN layers), a split window of 65 F compartments and an overlap of 17 compartments was adopted. The output of the Sub-Band Split block per channel has a format (B, T, 65,10).

[00159] Generally, for the examples above, but also for other implementations, an overlap between segments between 3 and 50 frequency compartments has proven beneficial.

[00160] The system has been reduced in order to present a Petition 870250093764, dated 10 / 14 / 2025, pp. 128 / 154 51 / 66 computational complexity of 24.93 MFlops with a minimal drop in subjective performance.

[00161] In this final section, the design choices related to multi-stage processing, the rationale for the choices, and the complexity reduction achieved are described. The Sub-Band Split per Channel method from the previous section reveals its limitation in reducing computational complexity. A smaller split window results in a larger number of channels; therefore, it is difficult for the main DNN to estimate the intermediate results.

[00162] Instead of using both the real and imaginary parts separately, the input is divided into features based on magnitude and phase. Regarding the noise reduction task, magnitude is subjectively more significant, so the Sub-Band Split and Mix method per channel is employed on magnitude and processed using a smaller DNNM, as illustrated in Figure 11. Since magnitude equals 1 dimension, the input feature dimension in the DNNm is (B, T, 257, 1). Because the input dimension has been reduced to 1, a smaller split window can be introduced.

[00163] In a practical implementation, two different DNNms for two different systems are used. This results in a feature dimension of (B, T, 48,8) after the Sub-Band Division block per channel. The DNNm employed for this system has a computational complexity of 4.62 MFlops with a parameter size of 0.684M. In another case, the DNNm employed has a computational complexity of 2.74 MFlops with a parameter size of 0.362M. In the latter system, a 40 F bin division window with an overlap of 16 bins was used, resulting in a feature dimension of (B, T, 40,10).

[00164] In both cases, an even smaller DNN, called DNNmp, was used to process the fused phase information at the output. Petition 870250093764, dated 10 / 14 / 2025, pp. 129 / 154 52 / 66 intermediate. The DNNmp exhibits a computational complexity of 1.73 MFlops with 3.4K parameters. This resulted in an overall computational complexity of 6.36 MFlops for the first system and 4.48 MFlops for the last system.

[00165] The overall performance drop after employing this method is still within an acceptable range, but the computational complexity has been reduced by almost 30 times. This is very significant and allows the algorithm to be run on embedded devices at a sampling rate of 16 kHz.

[00166] Subsequently, other embodiments of the invention, for example, for the second aspect, are summarized as examples, wherein the reference numbers in parentheses do not constitute any limitation to the general principle.

[00167] 1. Apparatus for processing an information signal, comprising: A feature extractor 100 is used to extract a set of features from the information signal, wherein the feature extractor 100 comprises: a raw feature calculator 110 for calculating the raw feature results, each raw feature result having at least two raw feature components; and a raw feature compressor 120 for performing a dynamic range compression of at least two raw feature components in order to obtain at least two compressed raw feature components for each raw feature result, wherein the feature set comprises the compressed raw feature components; and a signal processor 600 for processing the feature set in order to obtain the processed information signal, wherein the information signal comprises an audio signal, a signal of Petition 870250093764, dated 10 / 14 / 2025, pp. 130 / 154 53 / 66 image, or a radar signal.

[00168] 2. Apparatus, according to example 1, in which the gross resource calculator 110 is configured to calculate a complex value as the gross resource result, the gross resource result having a real part and an imaginary part, or a magnitude and a phase, as at least two gross resource components; and

[00169] Wherein the crude compressor 120 is configured to perform compression of the real part and the imaginary part, or of magnitude and phase, in order to obtain the compressed crude resource components.

[00170] 3. Apparatus, according to example 1 or 2, in which the gross resource compressor 120 is configured to apply 121 a first compression function to a first gross resource component, and to apply 122 a second compression function to a second gross resource component, the second compression function being different from the first compression function.

[00171] 4. Apparatus, according to one of the previous examples, in which performing the compression involves applying 121, 122 a power law compression using a power value, in which the power value is less than 1.

[00172] 5. Apparatus, according to example 4, in which the power value is different for each raw resource component, or in which the resource extractor 110 is configured to provide, for a specific information signal, or for a specific processing task, performed by the signal processor 600, one or more different numbers, or in which the resource extractor 100 is configured to provide, for a specific information signal, or for a specific processing task, performed by the signal processor 600, a Petition 870250093764, dated 10 / 14 / 2025, pp. 131 / 154 54 / 66 or more different numbers derived by specific training for the specific information signal, or specific processing task, performed by the 600 signal processor.

[00173] 6. Apparatus, according to one of the previous examples, in which the raw resource calculator 110 is configured to calculate the components of raw resources as absolute numbers and associated signs, and in which the raw resource compressor 120 is configured to apply a compression function to the absolute numbers and retain the signs of the corresponding components.

[00174] 7. Apparatus, according to one of the preceding examples, wherein the information signal is the audio signal, in which the raw resource calculator 110 is configured to perform a time and frequency decomposition in order to decompose the audio signal into a time and frequency representation, or wherein the information signal is the image signal, and wherein the raw resource calculator 110 comprises performing a spatial transformation to decompose the image signal into a spatial frequency representation.

[00175] 8. Device, according to one of the previous examples, in which the signal processor 600 comprises: a neural network processor 20, 300 to process the feature set in order to obtain a processed feature set; A 350 resource decompressor to perform a combination of decompression and compression performed by the 120 raw resource compressor in order to obtain a resulting resource set; and a 500 post-processor to post-process the resulting resource set in order to obtain the processed information signal. Petition 870250093764, dated 10 / 14 / 2025, pp. 132 / 154 55 / 66

[00176] 9. Apparatus, according to example 8, in which the neural network processor 300 is configured to divide 301, 302 the set of features into multiple segments and arrange the multiple segments along the channel dimension in order to optimize a number of channels from at least 2 to a number greater than two or to a number that is an integer multiple of 2, and in which the neural network processor 300 comprises a neural network configured to receive, as an input, the multiple segments as input channels and generate, as an output, a number of output channels, in which the neural network processor is configured to stack 401,402 the output channels in order to obtain a stacked set of features.

[00177] 10. Device, according to example 9, in which the neural network processor 300 is configured to perform a combination of features 400, from the stacked feature set, using an additional neural network 420 that has a lower complexity than the complexity of the neural network.

[00178] 11. Apparatus, according to example 10, in which the neural network processor 300 is configured to divide 301, 302 the feature set into overlapping segments, in which the stacked feature set has a larger dimension than the dimension of the feature set, and in which the additional neural network 420 is configured to reduce the dimension of the stacked feature set to the dimension of the feature set or to a dimension smaller than the dimension of the feature set.

[00179] 12. Device, according to one of the previous examples, in which the information signal is an audio signal; where the audio signal sample rate is less than 24 kHz, Petition 870250093764, dated 10 / 14 / 2025, pp. 133 / 154 56 / 66 wherein one dimension of the feature set in a frequency dimension is less than 500 and greater than 200 and the number of channels is equal to 2 for two raw feature components, wherein the number of segments is between two and four and the number of channels is between four and eight, and wherein an overlap of segments is between 30 and 90 frequency compartments.

[00180] Device, according to one of the previous examples, in which the signal processor comprises one or more neural networks to process the set of features in order to obtain the processed information signal.

[00181] 14. Apparatus, according to one of the preceding examples, wherein the signal processor 600 comprises a neural network processor 300 for processing the feature set comprising the compressed raw feature components, wherein the neural network processor 300 is configured to send the resulting feature set, wherein the apparatus comprises a feature decompressor for performing a decompression, which coincides with the compression performed by the raw feature compressor 120, in order to obtain a resulting decompressed feature set, wherein the resulting decompressed feature set is a time and frequency mask, and wherein the signal processor 600 comprises a mask processor for applying the time and frequency mask, as a spectral gain mask, to the feature set extracted by the raw feature calculator 110 and a representation of the processed information signal,or to calculate a processing filter from the time and frequency mask and to apply the processing filter to the audio signal or feature set, Petition 870250093764, dated 10 / 14 / 2025, pp. 134 / 154 57 / 66 calculated by the raw resource calculator 110 in order to obtain a representation of the processed information signal.

[00182] 15. Method of processing an information signal, comprising: Extracting a set of features from the information signal, including: Calculate the raw feature results, each raw feature result having at least two raw feature components; and perform dynamic range compression of at least two raw feature components in order to obtain at least two compressed raw feature components for each raw feature result, wherein the feature set comprises the compressed raw feature components; and process the feature set in order to obtain the processed information signal, wherein the information signal comprises an audio signal, an image signal, or a radar signal.

[00183] 16. Computer program to perform, when running on a computer or processor, the method of example 15.

[00184] Subsequently, other embodiments of the invention, for example, for the third aspect, are summarized as examples, wherein the reference numbers in parentheses do not constitute any limitation to the general principle.

[00185] 1. Apparatus for processing an information signal, comprising: A feature extractor 100 to extract a set of features from the information signal, wherein each feature within the set of features comprises at least two feature components, and wherein the set of features comprises a first subset with the first feature components, and a second Petition 870250093764, dated 10 / 14 / 2025, pages 135 / 154 58 / 66 subset with the second feature components; and a 300 neural network processor, comprising: A first neural network 340 is used to receive, as input, the first subset and to send a first processed subset; a 350 combiner to combine the first subset and the second processed subset in order to obtain a combined subset; and a second 360 neural network to receive, as an input, the combined subset and to send a processed combined output, wherein the processed combined output represents a processed information signal, or wherein the apparatus is configured to calculate the processed information signal using the processed combined output, and wherein the complexity of the first 340 neural network is greater than the complexity of the second neural network.

[00186] 2. Apparatus, according to example 1, wherein the information signal is an audio signal, wherein the feature extractor 100 comprises a time and frequency decomposer for computing a time and frequency domain representation of the audio signal, wherein the first subset comprises magnitude values ​​of the time and frequency representation and the second subset comprises phase values ​​of the time and frequency representation; or wherein a first feature component, among at least two feature components, is more important than a second feature component, of at least two feature components.

[00187] 3. Device, according to one of the previous examples, in Petition 870250093764, dated 10 / 14 / 2025, pp. 136 / 154 59 / 66 where the 350 combiner is configured to concatenate the first subset and the second processed subset along a channel direction, and where the first neural network is configured to process the first subset with a first number of channels equal to or greater than 1, and where the second 360 neural network is configured to process the combined subset with a second number of channels, where the second number of channels is greater than the first number of channels.

[00188] 4. Device, according to example 3, in which the second number of channels is greater than the first number of channels by 1.

[00189] 5. Apparatus, according to one of the previous examples, wherein the first subset comprises magnitude values, and wherein the second subset comprises phase values, and wherein the first processed subset comprises magnitude values, and wherein the combinator 350 is configured to compute, for each component of the first processed subset and the second processed subset, a real part component and an imaginary part component; and wherein the second neural network 360 is configured to receive, as the combined subset, a subset comprising the real part components and the imaginary part components.

[00190] 6. Apparatus, according to example 5, wherein the first subset comprises magnitude values ​​of the feature set, and wherein the feature set comprises the complex frequency compartment inputs,

[00191] Where the first neural network 340 is configured to compute the first processed subset, such that the first processed subset represents a time mask and Petition 870250093764, dated 10 / 14 / 2025, pp. 137 / 154 60 / 66 frequency possessing only the magnitude values ​​for the frequency compartments; and

[00192] Wherein the combiner 350 is configured to combine the magnitude mask values ​​of the frequency compartments with the phase values ​​of the complex frequency compartment inputs of corresponding frequency compartments, such that a magnitude mask value of a frequency compartment is combined with a phase value of the complex frequency compartment input of the same frequency compartment.

[00193] 7. Device, according to one of the previous examples, in which the neural network processor 300 is configured to: Perform a channel 301, 302 division of the first subset into M multiple segments, wherein the first neural network 340 is configured to receive the first subset concatenated along a channel direction with M channels and to send the first processed subset, and wherein the neural network processor 300 is configured to combine the first processed subset concatenated along the channel direction 420 into a single channel representation as the first processed subset.

[00194] 8. Apparatus, according to example 7, in which the multiple segments comprise two, three or more segments, in which the three or more segments are overlapping segments, and in which an overlap between the overlapping segments is greater than 3 frequency compartments and less than 50 frequency compartments.

[00195] 9. Device, according to example 7, in which the neural network processor 300 is configured to perform the division of Petition 870250093764, dated 10 / 14 / 2025, pp. 138 / 154 61 / 66 sub-band per channel 301,302 in M ​​overlapping segments of the first subset and wherein the first processed subset has M channels; and wherein the neural network processor 300 is configured to perform a sub-band per channel division of the second subset into M overlapping segments and arrange the second subset within the M channels, and wherein the combiner 350 is configured to perform processing of the first subset and the second subset, so that the combined subset comprises twice the number of M channels.

[00196] 10. Apparatus, according to example 9, in which the second neural network 360 is configured to process the combined subset that has twice the number of M channels in order to obtain an output from the second neural network 360, and in which the neural network process 300 is configured to stack 401, 402 and combine a set of stacked features within the processed combined output presenting a dimension similar to a dimension of the feature set extracted by the feature extractor 100.

[00197] 11. Device, according to one of the previous examples, where the complexities of the first 340 neural network and the second 360 neural network are measured in floating-point operations (FLOPS), where a higher number of floating-point operations represents greater complexity, or in execution time on one or more target hardware devices, where a higher execution time represents greater complexity, or in the power consumption of a given device, where higher power consumption represents greater complexity, or in the number of MAC (Multiplication and Accumulation) operations, where a higher number of operations Petition 870250093764, dated 10 / 14 / 2025, pp. 139 / 154 62 / 66 MAC represents greater complexity.

[00198] 12. Apparatus, according to one of the previous examples, in which the resource extractor 100 comprises: a gross resource calculator 110 for calculating the gross resource results, wherein each gross resource result has at least two gross resource components; and a gross resource compressor 120 for performing a compression of at least two gross resource components to obtain at least two gross resource components for each gross resource result, wherein the first subset comprises the first resources, among the at least two gross resource components, of the gross resource results, and wherein the second subset comprises the second gross resource components, among the at least two gross resource components, of the gross resource results.

[00199] 13. Device, according to one of the previous examples, in which the information signal comprises an audio signal, an image signal or a radar signal.

[00200] 14. Device, according to one of the previous examples, wherein the device is configured as an embedded device, or wherein the first 340 neural network and the second 360 neural network are configured to operate in series with each other.

[00201] 15. Method of processing an information signal, comprising: extract a set of features from the information signal, wherein each feature of the feature set comprises at least two feature components, and wherein the feature set comprises a first subset with the first feature components and a second subset with the second feature components; and Petition 870250093764, dated 10 / 14 / 2025, pp. 140 / 154 63 / 66 use a neural network processor 300 comprising: a first neural network 340 to receive, as an input, the first subset and to send a first processed subset; a 350 combiner to combine the first processed subset and the second processed subset in order to obtain a combined subset; and a second 360 neural network to receive, as an input, the combined subset and to send a processed combined output, wherein the processed combined output represents a processed information signal, or wherein the apparatus is configured to calculate the processed information signal using the processed combined output; and wherein the complexity of the first 340 neural network is greater than the complexity of the second neural network.

[00202] 16. Computer program to perform, when running on a computer or processor, the method according to example 15.

[00203] It should be understood that all alternatives or aspects discussed above and all aspects defined by the independent claims in the following claims may be used individually, that is, without any other alternative or objective besides the alternative, objective or independent claim contemplated. However, in other embodiments, two or more of the alternatives, aspects or independent claims may be combined with each other and, in other embodiments, all aspects, alternatives or independent claims may be combined with each other.

[00204] Although some aspects have been described in the context Petition 870250093764, dated 10 / 14 / 2025, pp. 141 / 154 64 / 66 of a device, it is clear that these aspects also represent a description of the corresponding method, where a block or device corresponds to a method step or a feature of a method step. Similarly, the aspects described in the context of a method step also represent a description of a corresponding block, item, or feature of a corresponding device.

[00205] Depending on certain implementation requirements, embodiments of the invention may be implemented in hardware or in software. Implementation may be carried out using a digital storage medium, for example, a floppy disk, a DVD, a CD, a ROM, a PROM, an EPROM, an EEPROM, or a FLASH memory, having stored electronically readable control signals that cooperate (or may cooperate) with a programmable computer system so that the respective method is carried out.

[00206] Some embodiments of the invention comprise a data carrier that has electronically readable control signals, which can cooperate with a programmable computer system, so that one of the methods described herein is carried out.

[00207] Generally, embodiments of the present invention can be implemented as a computer program product with program code, the program code operating to perform one of the methods when the computer program product runs on a computer. The program code can, for example, be stored in a machine-readable carrier.

[00208] Other embodiments comprise the computer program for performing one of the methods described herein, stored on a machine-readable carrier or storage medium. Petition 870250093764, dated 10 / 14 / 2025, pp. 142 / 154 65 / 66 is not transitional.

[00209] In other words, an embodiment of the inventive method is therefore a computer program that has program code to perform one of the methods described herein, when the computer program runs on a computer.

[00210] An additional embodiment of the inventive methods is, therefore, a data carrier (or a digital storage medium, or a computer-readable medium) comprising, recorded thereon, the computer program for carrying out one of the methods described herein.

[00211] An additional embodiment of the inventive method is, therefore, a sequence of data or a sequence of signals representing the computer program for carrying out one of the methods described herein. The sequence of data or the sequence of signals may, for example, be configured in order to be transferred via a data communication connection, for example, via the Internet.

[00212] An additional embodiment comprises a processing means, for example, a computer, or a programmable logic device, configured or adapted to perform one of the methods described herein.

[00213] An additional embodiment comprises a computer that has installed on it a computer program to perform one of the methods described herein.

[00214] In some embodiments, a programmable logic device (e.g., a field-programmable gate assembly) can be used to perform some or all of the functionalities of the methods described herein. In some embodiments, a field-programmable gate assembly can cooperate with a microprocessor in order to perform one of the methods described herein. Generally, the methods are preferably performed by any hardware device. Petition 870250093764, dated 10 / 14 / 2025, pp. 143 / 154 66 / 66

[00215] The embodiments mentioned above are merely illustrative of the principles of the present invention. It is understood that modifications and variations of the arrangements and details described herein will be apparent to persons skilled in the art. It is intended, however, that one be limited only by the scope of the patent claims and not by the specific details presented through description and explanation of the embodiments. References [1] Yu, G., Guan, Y., Meng, W., Zheng, C., & Wang, H. (2022). DMFNET: A decoupling-style multi-band fusion model for full-band speech enhancement. [2] Zhang, Xu, Lianwu Chen, Xiguang Zheng, Xinlei Ren, Chen Zhang, Liang Guo, and Bin Yu. “A Two-Step Backward Compatible Fullband Speech Enhancement System.” ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2022): 7762-7766. [3] S. Zhao, B. Ma, K. N. Watcharasupat e W. -S. Gan, “FRCRN: Boosting Feature Representation Using Frequency Recurrence for Monaural Speech Enhancement,” ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2022, pp. 9281-9285, doi: 10.1109 / ICASSP43922.2022.9747578. [4] Liu, Haohe, Lei Xie, Jian Wu e Geng Yang. “Channel-wise Subband Input for Better Voice and Accompaniment Separation on High Resolution Music.” ArXiv abs / 2008.05216 (2020): n.pag. [5] Lv, Shubo, Yihui Fu, Mengtao Xing, Jiayao Sun, Lei Xie, Jun Huang, Yannan Wang e Tao Yu. “S-DCCRN: Super Wide Band DCCRN with Learnable Complex Feature for Speech Enhancement.” ICASSP 2022 2022 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) (2022): 7767-7771. Petição 870250093764, de 14 / 10 / 2025, pág. 144 / 154

Claims

1 / 6 CLAIMS 1. Apparatus for processing an information signal, the information signal comprising an audio signal, an image signal or a radar signal, the apparatus being characterized in that it comprises: a feature extractor (100) for extracting a set of features from the information signal, wherein the feature extractor (100) comprises: a raw feature calculator (110) for calculating the raw feature results, each resulting raw feature having at least two raw feature components; and a raw feature compressor (120) for performing a dynamic range compression to at least two raw feature components in order to obtain at least two compressed raw feature components for each resulting raw feature, wherein the feature set comprises the compressed raw feature components;and a signal processor (600) to process the set of features in order to obtain the processed information signal, wherein the signal processor (600) comprises one or more neural networks.

2. Apparatus, according to claim 1, characterized in that the gross resource calculator (110) is configured to calculate a complex value as the resulting gross resource, the resulting gross resource having a real part and an imaginary part or a magnitude and a phase, as the at least two components of gross resources, and wherein the gross processor (120) is configured to perform a compression of the real part and the imaginary part, or of the magnitude and the phase, in order to obtain the compressed gross resource components. Petition 870250093764, dated 10 / 14 / 2025, pp. 145 / 154 2 / 6 3. Apparatus, according to claim 1 or 2, characterized in that the raw resource compressor (120) is configured to apply (121) a first compression function to a first raw resource component, and apply (122) a second compression function to a second raw resource component, the second compression function being different from the first compression function.

4. Apparatus, according to any of the preceding claims, characterized in that the embodiment of the compression comprises applying (121, 122) a power-law compression using a power value, wherein the power value is less than 1.

5. Apparatus, according to claim 4, characterized in that the power value is different for each component of raw resources, or in that the resource extractor (110) is configured to provide, for a specific information signal or for a specific processing task performed by the signal processor (600), one or more different numbers, or in that the resource extractor (100) is configured to provide, for a specific information signal or for a specific processing task performed by the signal processor (600), one or more different numbers being derived by specific training for the specific information signal or the specific processing task performed by the signal processor (600).

6. Apparatus, according to any of the preceding claims, characterized in that the gross resource calculator (110) is configured to calculate the components of gross resources as absolute numbers and the associated signs; and in that the gross resource compressor (120) is configured to apply a compression function to the absolute numbers and to retain the signs of the corresponding components.

7. Apparatus, according to any of the preceding claims, characterized in that the information signal is an audio signal, wherein the raw feature calculator (110) is configured to perform a time and frequency decomposition in order to decompose the audio signal into a time and frequency representation, or wherein the information signal is the image signal, and wherein the raw feature calculator (110) comprises performing a spatial transformation in order to decompose the image signal into a spatial frequency representation.

8. Apparatus, according to any of the preceding claims, characterized in that the signal processor (600) comprises: a neural network processor (20, 300) comprising one or more neural networks, wherein the neural network processor (20, 300) is configured to process the feature set in order to obtain a processed feature set; a feature decompressor (350) to perform a combination of decompression and compression performed by the raw feature compressor (120) in order to obtain a resulting feature set; and a post-processor (500) to post-process the resulting feature set in order to obtain the processed information signal.

9. Apparatus, according to claim 8, characterized in that the neural network processor (300) is configured to divide (301,302) the set of features into multiple segments and arrange the multiple segments along the channel dimension in order to refine a number of channels from at least 2 up to a number greater than two or a number that is an integer multiple of 2, and Petition 870250093764, of 10 / 14 / 2025, pp. 147 / 154 4 / 6 wherein the neural network processor (300) comprises a neural network configured to receive, as an input, the multiple segments as input channels and to generate, as an output, a number of output channels, wherein the neural network processor is configured to stack (401,402) the output channels in order to obtain a stacked set of features.

10. Apparatus, according to claim 9, characterized in that the neural network processor (300) is configured to perform a combination of features (400) from the stacked feature set using an additional neural network (420) that has a lower complexity than the complexity of the neural network.

11. Apparatus, according to claim 10, characterized in that the neural network processor (300) is configured to divide (301, 302) the feature set into overlapping segments, wherein the stacked feature set has a larger dimension than the dimension of the feature set, and wherein the additional neural network (420) is configured to reduce the dimension of the stacked feature set to the dimension of the feature set or to a dimension smaller than the dimension of the feature set.

12. Apparatus, according to any of the preceding claims, characterized in that the information signal is an audio signal; wherein a sample rate of the audio signal is less than 24 kHz; wherein a dimension of the feature set in a frequency dimension is less than 500 and greater than 200 and a number of channels is equal to 2 for the two raw feature components; wherein the number of segments is between two and four and the number of channels is between four and eight; and wherein an overlap of the segments is between 30 and 90 frequency compartments.

13. Device, according to any of the preceding claims, characterized in that: the signal processor comprises one or more neural networks to process the set of features in order to obtain the processed information signal.

14. Apparatus, according to any of the preceding claims, characterized in that: the signal processor (600) comprises a neural network processor (300) for processing the feature set comprising the compressed raw feature components, wherein the neural network processor (300) is configured to send the resulting feature set; wherein the apparatus comprises a feature decompressor for performing a combination of decompression and compression performed by the raw feature compressor (120) in order to obtain a resulting decompressed feature set; wherein the resulting decompressed feature set is a time and frequency mask;and wherein the signal processor (600) comprises a mask processor for applying the time and frequency mask as a spectral gain mask to the feature set extracted by the raw feature calculator (110) and a representation of the processed information signal, or for calculating a processing filter from the time and frequency mask and for applying the processing filter to the audio signal or to the feature set calculated by the raw feature calculator (110) in order to obtain a representation of the processed information signal. Petition 870250093764, of 10 / 14 / 2025, pp. 149 / 154 6 / 6; 15. Method for processing an information signal, the information signal comprising an audio signal, an image signal or a radar signal, the method being characterized in that it comprises: extracting a set of features from the information signal, comprising: calculating raw feature results, each resulting raw feature having at least two raw feature components; and performing a dynamic range compression of the at least two raw feature components in order to obtain at least two compressed raw feature components for each resulting raw feature, wherein the feature set comprises the raw feature components; and processing the feature set in order to obtain the processed information signal, wherein the processing comprises using one or more neural networks.

16. Computer-readable medium characterized in that it comprises instructions for performing, when executed on a computer or processor, the method as defined in claim 15. Petition 870250093764, dated 10 / 14 / 2025, pp. 150 / 154

Images