Device and method for improving voice synthesis produced by an existing vocoder by means of a predetermined codec, and computer program therefor
Patent Information
- Application Number
- AE202602966
- Authority / Receiving Office
- AE · AE
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2024-03-05
- Filing Date
- 2025-03-05
Smart Images
Figure ABST_ABST
Abstract
Description
TITLE: Device and method for improving voice synthesis produced by an existing vocoder by means of a predetermined codec, and computer program therefor The present invention relates to a device for improving the speech synthesis produced, by means of a predetermined codec, by an existing vocoder.The present invention also relates to a method for improving the speech synthesis produced, by means of a predetermined codec, by an existing vocoder.The invention also relates to a computer program comprising software instructions which, when implemented by a computer, implement such a method. The invention relates to the field of speech coding also called “vocoding”, speech synthesis and voice transmission.Hereinafter, according to the present invention, “vocoding” and, by extension, the corresponding device “vocoder” are associated with the contraction of the English words Voice Coding and Voice coder, respectively, in other words vocal coding by an electronic device for processing the sound signal, substantially corresponding to a method according to which, from a decomposition into important properties (e.g. main spectral components) of an input voice or of another sound, an associated synthetic voice is synthesized at the output.Digital communications have a significant number of benefits compared to analog communications. In addition to being of better quality, they are often more robust to interfering noise during transmission. However, one of the major problems related to the very nature of digital communications is the limitation of the bit rate. Indeed, the higher the bit rate, the more it is possible to send a large amount of information allowing a higher voice quality. In addition, digital transmissions suffer from a “binarization” of the data received. Indeed, either the transmission of the digital frames is correct and the communication is unchanged, or it is incorrect (i.e. missing frames) and all of the information contained in these frames missing upon reception is lost.Certain transmission contexts impose constrained bit rates and the use of codecs (or coder-decoder from the English coder-decoder) whose sound quality is limited. In these contexts, the codecs used are generally standardized such as the MELP codec (from the English Mixed-excitation linear prediction) or else the enhanced MELPe codec (from the English enhanced Mixed-excitation linear prediction) most often associated with a bit rate, for example MELP 1200 corresponding to a MELP codec with a bit rate of 1200 bits / s.The codecs are generally optimized to operate according to a precise modeling (for example LPC (from the English Linear Predictive Coder, MELP) as regards the extraction of the parameters of the speech signal.The parametric model obtained can be used to synthesize a speech signal. This is referred to as a parametric coder or a vocoder (to specify that the model applies to the speech signal).The quantization / dequantization step allows transmission over a constrained channel by restricting the bit rate. There are therefore two sources of degradation: approximation of the model, and quantization of the parameters of the model.The different bit rates of MELP correspond to different operating modes. It is the quantization scheme that is optimized for a given model (LPC, MELP, …) and a target bit rate. If the binary streams are obviously different, the information is of the same type (associated with the MELP parameters).A binary stream associated with the MELPe 1200 codec operates with the MELPe 1200 vocoder and cannot operate with an MELPe 2400 vocoder.Indeed, a modification of the quantization scheme implies a loss of interoperability: an MELP 1200 is not interoperable with an MELP 2400.Moreover, with these conventional low-bit-rate solutions, a loss of frame transmission is accompanied by an interruption during decoding of the communication.The aim of the invention is then to improve the sound quality and to extend the robustness of the low-bit-rate or even very-low-bit-rate codecs used in certain transmission contexts which impose constrained bit rates, such as for example the aforementioned MELPe 1200 codec, while remaining interoperable with the existing vocoders.For this purpose, the invention has as its subject device for improving the speech synthesis produced, by means of a predetermined codec, by an existing vocoder, said device comprising:- a parameter reception module suitable for receiving the parameters characterizing each frame of the speech synthesis capable of being produced by said existing vocoder;- at least one speech synthesizer with artificial intelligence having an architecture according to a neural source-filter model, said speech synthesizer with artificial intelligence comprising at least:- a projection module configured to project, via a first unidirectional recurrent neural network, said parameters, or parameters derived from said parameters, into a latent space of parameterizable dimension;- a harmonic signal generator configured to generate, via a second neural network, a signal representative of the harmonics of said speech synthesis associated with said parameters received via said reception module;- a speech synthesis module configured to synthesize, via a third unidirectional recurrent neural network, an encoded voice associated with said human voice from:- the projection of said parameters projected into said latent space of parameterizable dimension provided by said projection module, and- of said signal representative of the harmonics of the speech synthesis associated with said parameters received via said reception module and provided by said generator;said speech synthesizer with artificial intelligence having been previously trained during a supervised learning phase using training speech syntheses produced by means of said predetermined codec.It should be noted that by “improvement” is meant a better perception of the quality of the transmitted voice but also better robustness to transmission interruptions (i.e. frame losses).The improvement device according to the present invention is advantageous, because it proposes using a speech synthesizer with artificial intelligence to improve the speech synthesis of an already existing vocoder.The AI artificial intelligence speech synthesizer of the improvement device according to the present invention is advantageous because, unlike neural source-filter architectures (from the English source-filter) of the literature, the use of unidirectional recurrent neural networks compatible with real-time use, that is to say with a latency of less than 50ms.Furthermore, the speech synthesizer with artificial intelligence according to the present invention is configured to provide a frugal speech synthesis solution, because it is based on a data-frugal architecture. Indeed, the source-filter model is admittedly a conventional model for decomposing the voice into source and filter(s) components, however the use of this model, boosted by “modern” unidirectional recurrent neural networks (i.e. suitable for restoring causality) makes it possible to offer a framework conducive to voice modeling. Such “boosting by artificial intelligence” results in better reconstruction of the voice with a lower memory and computation cost.The AI artificial intelligence speech synthesizer according to the present invention has the advantage of optimally learning the exploitation of the input data, namely the parameters provided by an existing vocoder, to accomplish its task, that is to say the restoration of the voice communication. To do this, said speech synthesizer with artificial intelligence of the proposed improvement device is specialized in an operating mode after a so-called supervised learning phase where said AI artificial intelligence speech synthesizer is trained from a binary stream of a low-bit-rate codec. Such prior training then allows, during the inference phase, a significant improvement in the quality of the result originally provided by the low-bit-rate codec.More precisely, the neural network does not use the binary stream directly here but the parameters of the model previously reconstructed from the binary stream (dequantization step) or even other parameters derived from the parameters of the model (as a replacement for or in addition to the latter).According to other advantageous aspects of the invention, the improvement device comprises one or more of the following features, taken in isolation or according to all the technically possible combinations:- the reception module comprises an identification tool for said predetermined codec by analyzing the speech synthesis produced by said existing vocoder, said identification tool being further configured to select, from among a plurality of speech synthesizers with artificial intelligence previously trained and respectively associated with distinct codecs, said speech synthesizer with artificial intelligence corresponding to said identified predetermined codec;- the reception module comprises an interpolation tool configured to linearly interpolate said parameters and provide as input to said artificial intelligence speech synthesizer interpolated parameters compatible with said speech synthesis module;- said signal generated by the generator and representative of the harmonics of said speech synthesis associated with said parameters received via said reception module is a weighted sum of the signal corresponding to the pitch associated with said received parameters and its harmonics, said weighting having been learned beforehand by said second neural network;- white noise is added to said weighted sum;- the artificial intelligence speech synthesizer further comprises an adversarial training module comprising two discriminators configured to be trained simultaneously with said artificial intelligence speech synthesizer during its supervised learning phase, according to a zero-sum game, said discriminators being trained to differentiate real signals from those suitable for being synthesized by said artificial intelligence speech synthesizer, said artificial intelligence speech synthesizer being trained to deceive said two discriminators;- one of said two discriminators is sensitive to low-frequency information and the other of said discriminators is sensitive to high-frequency information;- the artificial intelligence speech synthesizer further comprises:- a reconstruction module configured to reconstruct missing frames of said synthesized encoded voice from the frames preceding or following said missing frames, and / or- a band extension module configured to reconstruct a signal comprising said synthesized encoded voice with a sampling frequency higher than that of said speech synthesis associated with said parameters received via said reception module.The invention also relates to a method for improving the speech synthesis produced, by means of a predetermined codec, by an existing vocoder, said method being implemented by an improvement device as previously described, said method comprising:- receiving the parameters characterizing each frame of the speech synthesis suitable for being produced by said existing vocoder;and the following steps implemented during an inference phase by an artificial intelligence speech synthesizer of said improvement device, said artificial intelligence speech synthesizer having been previously trained during a supervised learning phase using training parameters characterizing each frame of speech syntheses suitable for being produced by means of said predetermined codec:- projecting, via a first unidirectional recurrent neural network of said artificial intelligence speech synthesizer, the parameters received via said reception module or parameters derived from said parameters, into a latent space of parameterizable dimension;- generating, via a second neural network of said artificial intelligence speech synthesizer, a signal representative of the harmonics of the speech synthesis associated with said parameters received via said reception module;- speech synthesis, via a third unidirectional recurrent neural network, of an encoded voice associated with said human voice from:- the projection of said parameters projected into said latent space of parameterizable dimension provided by said projection module, and- said signal representative of the harmonics of the speech synthesis associated with said parameters received via said reception module and provided by the generator of the artificial intelligence speech synthesizer of said improvement device.The invention also relates to a computer program comprising software instructions which, when executed by a computer, implement a speech synthesis method, as defined above. The invention will appear more clearly upon reading the following description, given solely by way of non-limiting example, and made with reference to the drawings in which:[Fig. 1] figure 1 is a general schematic representation of a device for improving the speech synthesis produced, by means of a predetermined codec, by an existing vocoder according to the present invention;[Fig. 2] figure 2 illustrates the use of the improvement device of figure 1;[Fig. 3] figure 3 illustrates an example of an architecture of the improvement device of figure 1;[Fig. 4] figure 4 illustrates the implementation of an optional module of the improvement device of figure 1 during its training phase;[Fig. 5] figure 5 is a flowchart of a method for improving the speech synthesis produced, by means of a predetermined codec, by an existing vocoder implemented by the improvement device of figure 1 according to an embodiment. In the remainder of the description, the expression “substantially equal to” is understood as an equality relationship to within plus or minus 10%, that is to say with a variation of at most 10%, preferably still as an equality relationship to within plus or minus 5%, that is to say with a variation of at most 5%.Furthermore, thereafter, it is considered that a neural network comprises an ordered succession of neuron layers, each of which takes its inputs from the outputs of the preceding layer.More precisely, each layer comprises neurons taking their inputs from the outputs of the neurons of the preceding layer, or from the input variables for the first layer.As a variant, more complex neural network structures may be contemplated with a layer which may be connected to a layer more remote than the immediately preceding layer.Each neuron is also associated with an operation, that is to say a type of processing, to be performed by said neuron within the corresponding processing layer.Each layer is connected to the other layers by a plurality of synapses. A synaptic weight is associated with each synapse, and each synapse forms a link between two neurons. It is often a real number, which takes positive as well as negative values. In some cases, the synaptic weight is a complex number.Each neuron is suitable for performing a weighted sum of the value(s) received from the neurons of the preceding layer, each value then being multiplied by the respective synaptic weight of each synapse, or link, between said neuron and the neurons of the preceding layer, and then for applying an activation function, typically a non-linear function, to said weighted sum, and for delivering at the output of said neuron, in particular to the neurons of the following layer which are connected thereto, the value resulting from the application of the activation function. The activation function makes it possible to introduce a non-linearity into the processing performed by each neuron. The sigmoid function, the hyperbolic tangent function, and the Heaviside function are examples of activation functions.As an optional addition, each neuron is also capable of further applying a multiplicative factor, also called a bias, to the output of the activation function, and the value delivered at the output of said neuron is then the product of the bias value and the value resulting from the activation function.A fully connected layer of neurons is a layer in which the neurons of said layer are each connected to all the neurons of the preceding layer.Such a type of layer is more often referred to by the English term « fully connected », and is sometimes designated by the name « dense layer ».The electronic device for improving the produced speech synthesis, by means of a predetermined codec, by an existing vocoder is illustrated in Figure 1.Such an existing vocoder classically produces a parametric model of the speech signal with, on the one hand, an analysis module implementing the extraction of the parameters and, on the other hand, a synthesis module implementing the reconstruction of the signal from the extracted parameters.By codec is meant the combination of an encoder and a decoder (i.e. encoder + decoder), the encoder for its part comprising an analysis module and a module for quantizing the parameters resulting from the analysis, the decoder being suitable for carrying out the dequantization (i.e. the restoration of the parameters from the bitstream) and then the speech synthesis as such.The electronic device for improving the speech synthesis is located at the output of the dequantization module of the existing vocoder (i.e. conventional, such as a MELP 1200 vocoder), that is to say after the dequantization module of the decoder of the codec conventionally configured to dequantize (i.e. restore the parameters from the bitstream).In other words, the electronic device for improving the speech synthesis according to the present invention uses as input the parameters SV resulting from the dequantization of the existing vocoder and / or parameters derived from the latter, and replaces the pre-existing synthesis module of the existing vocoder. Thus, the synthesis model of the existing vocoder is no longer used but is replaced by the speech synthesizer 10 described below which uses as input the parameters resulting from the dequantization of the existing vocoder 200 and / or parameters derived from the latter.Such an electronic device for improving the speech synthesis first of all comprises a parameter reception module characterizing each frame of the speech synthesis suitable for being produced by said existing vocoder.An example of an architecture of this reception module 110 is in particular described below in relation to Figure 3.As an option, the reception module 110 comprises an identification tool for said predetermined codec, the identification tool being configured to determine said predetermined codec by analyzing the parameters characterizing each frame of the speech synthesis suitable for being produced by said existing vocoder.Furthermore, said optional identification tool is also configured to select, from among a plurality, not shown, of artificial-intelligence speech synthesizers previously trained and respectively associated with distinct codecs, said artificial-intelligence speech synthesizer 10 corresponding to said identified predetermined codec.More precisely, the reception module 110 is configured at a given instant to operate with a speech encoder (there is a negotiation between the transmitter and the receiver to use a common encoder). On reception, the information is therefore available and the previously trained and adapted artificial-intelligence speech synthesizer can be selected or configured (for example, same network topology but different weights for a different quantization scheme but an identical parametric model) from this information.In other words, such an identification tool makes it possible to adapt the artificial-intelligence speech synthesizer to be used subsequently, as a function of the received bitstream which it used to identify the codec used by the existing vocoder whose speech synthesis is to be improved according to the present invention. Thus, the identification tool is suitable for implementing a phase of identifying the codec used by the existing vocoder in order to then switch to the artificial-intelligence speech synthesizer 10 adapted to the received parameters SV, in particular when the improvement device comprises beforehand a plurality of artificial-intelligence speech synthesizers previously trained and respectively associated with distinct codecs.Thus, said selected artificial-intelligence speech synthesizer 10 was previously trained during a supervised learning phase using training speech syntheses produced by means of said identified predetermined codec. In other words, supervised training adapted to the use case is implemented by exploiting the parameters encoded by the codec of the existing vocoder whose speech synthesis it is desired to improve.According to another variant, the reception module 110 does not comprise an identification tool, and the artificial-intelligence speech synthesizer 10 is directly adapted to the existing vocoder, namely it was previously trained with training speech syntheses produced by means of the predetermined codec of the existing vocoder, for example by means of an MELPe 1200 bitstream making it possible to guarantee the interoperability condition with an existing MELPe 1200 vocoder. Once trained, the artificial-intelligence speech synthesizer 10 therefore receives, via its reception module 110, as input the current decoded parameters provided by the decoder of the existing MELPe 1200 vocoder.As can be seen in this Figure 1, the artificial-intelligence speech synthesizer 10 comprises a projection module 12 configured to project, via a first unidirectional recurrent neural network RN1, parameters characterizing each frame of said speech synthesis suitable for being produced by the existing vocoder, received via said reception module, into a latent space (in English embedding) of parameterizable dimension.The parameters at the input of the unidirectional recurrent network RN1 may be the parameters of the encoder used of the existing vocoder or may be replaced or supplemented by parameters obtained by transformation of the initial parameters.For example, for MELP, the parameters characterizing the spectral envelope of the speech signal are the LSF parameters (from the English Line Spectral Frequencies) directly associated with the LPC model. These parameters may be used as input to the unidirectional recurrent network RN1 or may be replaced by other parameters such as: cepstral coefficients (LPCC from the English Linear Prediction Ceptral Coefficients) or log-spectral coefficients obtained by DCT transformation (from the English: Discrete Cosine Transform) of the cepstral coefficients. These transformations may make it possible to exploit orthogonalization properties or better resolution by increasing the number of parameters (for example ten LSF coefficients, eighteen LPCC coefficients, sixty-four spectral coefficients, etc.).The LPCC coefficients are obtained by transformation of the linear prediction coefficients (LPC model). The conversion relationships between the LSF coefficients, the linear prediction coefficients and the LPCC coefficients are established and known in the field of speech processing.Such a projection module 12 is also called an encoder of the artificial-intelligence speech synthesizer 10, and is therefore suitable for “embedding” the parameters characterizing each frame of the speech synthesis suitable for being produced by said existing vocoder and received via said reception module 110 into a latent space (i.e. a vector space) of parameterizable dimension, for example of a size equal to 128 (i.e. capable of comprising 128 components). The neural network RN1, previously trained as will be seen below, handles the nonlinear projection of said parameters into the latent space potentially containing more information than the parameters since temporal relationships may be incorporated therein owing to the recurrent and unidirectional nature of the neural network RN1.The parameters suitable for being handled by said projection module 12 are parameters characterizing each frame of the speech synthesis suitable for being produced by said existing vocoder and received via said reception module 110. The nature of these parameters is imposed by the existing vocoder.As illustrated by Figure 1, the artificial-intelligence speech synthesizer 10 further comprises a harmonic signal generator 14 configured to generate, via a second neural network RN2, a signal representative of the harmonics of said speech synthesis associated with said parameters received via said reception module 110.As an optional addition, such a harmonic signal generator 14 is configured to generate said signal representative of the harmonics of said speech synthesis associated with said parameters received via said reception module 110 which is a weighted sum of the signal corresponding to the pitch associated with said speech synthesis itself associated with said received parameters and of its harmonics, said weighting being learned beforehand by said second neural network.According to a variant of this optional addition, white noise is added to said weighted sum.In other words, said harmonic signal generator 14 is configured to produce a signal composed of a sinusoidal signal at the pitch associated with said speech synthesis associated with said received parameters and of its harmonics. The signal at the output of this generator 14 is a weighted sum of the pitch associated with said speech synthesis associated with said received parameters and of its harmonics. The weighting is learned by the second neural network RN2 during a prior training phase. White noise is added to this harmonic signal to serve as a basis for the stochastic components of the voice. The unvoiced frames or frames containing no voice are composed only of white noise without a harmonic component.According to a speech synthesizer architecture according to the known “source-filter” model, the output generated by the harmonic signal generator 14 plays the role of the “source” part of this “source-filter” architecture.As illustrated by Figure 1, the artificial-intelligence speech synthesizer 10 further comprises a speech synthesis module 16 configured to synthesize, via a third unidirectional recurrent neural network RN3, an encoded voice (associated with the human voice of the speaker having produced a sound signal at the input of the existing vocoder) from:- the projection of said parameters projected into said latent space of parameterizable dimension provided by said projection module 12, and- said signal representative of the harmonics of said speech synthesis associated with said parameters received via said reception module 110 and provided by said generator 14.Such a speech synthesis module 16 is also called the decoder of the artificial-intelligence speech synthesizer 10, and like the preceding elements 12 and 14 of the artificial-intelligence speech synthesizer 10 is also based on a recurrent neural network architecture. Such a speech synthesis module 16 therefore has as inputs the output of the projection model 12 and the output of the harmonic signal generator 14. The projection model 12 and the harmonic signal generator 14 are suitable for operating in parallel or successively in any order, provided that the output of the projection model 12 and the output of the harmonic signal generator 14 associated with one and the same frame of said speech synthesis associated with said parameters received via said reception module 110 are processed simultaneously by said speech synthesis module 16.According to an architecture of an artificial-intelligence speech synthesizer 10 according to the known “source-filter” model, the speech synthesis module 16 plays the role of the “filter” part of this “source-filter” architecture.Thus, unlike the neural source-filter architectures according to the current state of the art, the use of unidirectional recurrent neural networks proposed according to the present invention makes it possible to operate in real time.For example, the artificial-intelligence speech synthesizer 10 according to the present invention is suitable for operating with said speech synthesis associated with said parameters received via said reception module 110 previously decomposed into 40 ms frames with a 40 ms anticipation buffer (in English look ahead), essentially due to the extraction of the parameters at the input of the projection module 12 (i.e. of the encoder).As an optional addition, as represented by dotted lines, said artificial-intelligence speech synthesizer 10 further comprises an adversarial training module 18 comprising two discriminators D1 and D2 configured to be trained simultaneously with said artificial-intelligence speech synthesizer 10 during its supervised learning phase, according to a zero-sum game, said discriminators D1 and D2 being trained to differentiate real signals from those suitable for being synthesized by said artificial-intelligence speech synthesizer 10, said artificial-intelligence speech synthesizer 10 being trained to fool said two discriminators D1 and D2.It should be noted that once the learning phase has been carried out, such an adversarial training module 18 is no longer used during the inference phase using the artificial-intelligence speech synthesizer 10 thus previously trained adversarially.According to an advantageous variant of this optional addition, one of said two discriminators is sensitive to low-frequency information and the other of said two discriminators is sensitive to high-frequency information.In other words, the two discriminators D1 and D2 are trained at the same time as the artificial-intelligence speech synthesizer 10 to differentiate real signals from signals synthesized by the artificial-intelligence speech synthesizer 10 by creating a zero-sum game where the discriminators D1 and D2 try to differentiate the signals created by the artificial-intelligence speech synthesizer 10 from the original signals and where the artificial-intelligence speech synthesizer 10 is trained to fool the two discriminators D1 and D2.Such a zero-sum game makes it possible to balance the artificial-intelligence speech synthesizer 10 resulting from said adversarial training, and the use of two discriminators, sensitive to low- and high-frequency information, allows the artificial-intelligence speech synthesizer 10 to correct its defects over extended frequency ranges, and is particularly effective for removing the metallic artifacts from the voice synthesized via the artificial-intelligence speech synthesizer 10.For example, the two discriminators correspond to the discriminators D1 and D2 described in the article by J. Kong et al. entitled “HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis” NeurIPS 2020, but applied in another context, without combined implementation of a speech synthesizer architecture according to a source-filter model and of the zero-sum game according to the present invention where said discriminators D1 and D2 are trained to differentiate real signals from those suitable for being synthesized by said artificial-intelligence speech synthesizer 10, said artificial-intelligence speech synthesizer 10 being trained to fool said discriminators D1 and D2.The adversarial training implemented via said optional adversarial training module 18 is further described hereinafter in relation to figure 4.Such an optional adversarial training module 18 therefore makes it possible to improve the synthesis quality.As an optional addition, said artificial-intelligence speech synthesizer 10 further comprises a reconstruction module 20 configured to reconstruct (in English inpainting) missing frames of said synthesized encoded voice from the frames preceding or following said missing frames, and / or a band extension module 22 configured to reconstruct a signal comprising said synthesized encoded voice with a sampling frequency higher than that of said speech synthesis associated with said parameters received via said reception module 110.The optional reconstruction module 20 makes it possible to improve the overall transmission of the signal synthesized via said artificial-intelligence speech synthesizer 10 by compensating for the micro-cuts that would occur without its implementation due to missing frame(s), whereas the optional band extension module 22 makes it possible to increase the output sampling frequency, for example to predict the missing information (between 4 and 8 kHz, for example, for a change from a sampling frequency of 8 to 16 kHz).As an optional addition, at least one of said neural networks of said artificial-intelligence speech synthesizer 10, for example the first unidirectional recurrent neural network RN1 of the projection module 12 or alternatively the second neural network RN2, is also configured to extract the vocal biomarkers of the original speaker having generated said speech synthesis associated with said parameters received via said reception module 110, after having been previously trained for this purpose in a supervised manner.In a non-represented variant, the artificial-intelligence speech synthesizer 10 is also suitable for optionally comprising another neural network dedicated to extracting vocal biomarkers of the speaker having generated said speech synthesis associated with said parameters received via said reception module 110 after having been previously trained for this purpose in a supervised manner.In the example of figure 1, the artificial-intelligence speech synthesizer 10 comprises an information processing unit 26 formed, for example, by a memory 28 and a processor 30 associated with the memory 28.In the example of figure 1, the projection module 12, the generator 14 and the speech synthesis module 16, as well as, as an optional addition, the adversarial training module 18, the reconstruction module 20 and the band extension module 22 are each produced in the form of software, or of a software block, executable by the processor 30. The memory 28 of the artificial-intelligence speech synthesizer 10 is then suitable for storing projection software, harmonicsignal generation software, and speech synthesis software, as well as, as an optional addition, adversarial training software, reconstruction software, band extension software. The processor is then suitable for executing each of the software items among the projection software, the harmonic signal generation software, and the speech synthesis software, as well as, as an optional addition, the adversarial training software, the reconstruction software, the band extension software.In a non-represented variant, the projection module 12, the generator 14 and the speech synthesis module 16, as well as, as an optional addition, the adversarial training module 18, the reconstruction module 20 and the band extension module 22 are each produced in the form of a programmable logic component, such as an FPGA (in English Field Programmable Gate Array), or alternatively of an integrated circuit, such as an ASIC (in English Application Specific Integrated Circuit).When the artificial-intelligence speech synthesizer 10 is produced in the form of one or more software items, that is to say in the form of a computer program, also called a computer program product, it is further suitable for being recorded on a medium, not represented, readable by a computer. The computer-readable medium is, for example, a medium suitable for storing electronic instructions and for being coupled to a bus of a computer system. By way of example, the readable medium is an optical disc, a magneto-optical disc, a ROM memory, a RAM memory, any type of non-volatile memory (for example FLASH or NVRAM) or a magnetic card. A computer program comprising software instructions is then stored on the readable medium.Figure 2 illustrates the use 32 of the artificial-intelligence speech synthesizer 10 of figure 1. According to this use, a speaker 34 first produces a sound signal 36 comprising his own human voice. Such a sound signal 36 is then transmitted to the input of the existing vocoder.The existing vocoder is configured to extract from the sound signal 36 the parameters conforming to the codec that it uses, to implement quantization and dequantization providing the decoded parameters 39 at the input of the speech synthesis enhancement device.The nature of these parameters 39 is imposed by the existing vocoder. These parameters 39 are provided at the input of the projection module 40 (corresponding to the projection module 12 of figure 2) which comprises the first unidirectional recurrent neural network RN1 configured for said parameters in a latent space of configurable dimension.In parallel, the decoded parameters 39 received via said reception module 110 are provided at the input of the harmonic signal generator 42 (corresponding to the generator 14 of figure 1), said generator 42 providing at the output a signal 44 representative of the harmonics of said speech synthesis associated with said parameters received via said reception module 110.Said signal 44 representative of the harmonics of said speech synthesis associated with said parameters received via said reception module 110 as well as the result provided by projection module 40 are associated with the same frame of said speech synthesis associated with said parameters received via said reception module 110 and simultaneously provided as input to speech synthesizer 46 configured to synthesize, via a third unidirectional recurrent neural network RN3, an encoded voice 48 associated with decoded parameters 39 received via said reception module 110 of the speech synthesis enhancement device.Figure 3 illustrates an example of an architecture of speech synthesis enhancement device 100. According to this example, speech synthesis enhancement device 100 receives as input parameters Sv characterizing each frame of the speech synthesis suitable for being produced by an existing vocoder, for which its own speech synthesizer with artificial intelligence 10 has been previously trained.For example, the existing vocoder is of the MELPe 1200 type (or else of the MELPe 2400 type, etc.) and provides a MELPe 1200 binary stream (respectively MELPe 2400). Speech synthesizer with artificial intelligence 10 has been previously trained with acoustic parameters derived from this type of MELPe 1200 binary stream (respectively MELPe 2400) in order to guarantee the interoperability condition with existing vocoder 200 of the MELPe 1200 type (respectively MELPe 2400). MELP 2400 uses a quantization scheme over a frame of 54 bits and duration 22.5 ms, MELP 1200 uses a quantization scheme over a “super-frame” composed of three consecutive frames = 3 x 22.5 ms = 67.5 ms. The quantization scheme makes it possible to exploit the redundancy of information and to reduce the bit rate to 81 bits / superframe.Reception module 110 therefore receives as input parameters Sv comprising, according to an example, the decoded parameters of MELPe 1200.As an optional addition, reception module 110 comprises an interpolation tool configured to linearly interpolate the received parameters and provide as input to said speech synthesizer with artificial intelligence 10 interpolated parameters compatible with said speech synthesis module 16.In other words, if the existing vocoder uses a frame duration unsuitable for speech synthesizer with artificial intelligence 10 of enhancement device 100, the parameters originating from the existing vocoder are interpolated so as to be synchronous with speech synthesis module 16 of AI speech synthesizer 10. Optionally, a change of parametric space is also applied.It is thus possible to temporally interpolate the parameters at the input of speech synthesizer with artificial intelligence 10 in order to adapt the operating rate of the network (often to increase the synthesis rate: for example from 22.5 ms to 10 ms). The interpolation may be carried out linearly (for example at the level of the LSF coefficients which have good interpolation properties, and which may be used via successive transformations from LSF to LPCC, then from LPCC to spectral coefficients, the transition relationships between the LSF coefficients, the linear prediction coefficients and the LPCC coefficients being established and known in the field of speech processing.Thus, optionally interpolated, the decoded parameters of the existing vocoder are extracted, namely for example the variables corresponding in particular to the linear prediction cepstral coefficients LPCC (in English linear prediction cepstral coefficients), to the dynamic linear prediction cepstral coefficients DLPCC (in English dynamiclinear prediction cepstral coefficients), to the gain and to the voicing.According to the example of the architecture of figure 3, these parameters are preprocessed upstream of speech synthesizer with artificial intelligence 10.More precisely, the variables corresponding to the linear prediction cepstral coefficients LPCC (in English linear prediction cepstral coefficients), and to the dynamic linear prediction cepstral coefficients DLPCC (in English dynamiclinear prediction cepstral coefficients) are processed by a normalization tool configured to center and normalize them (i.e. normalize them on the scale of the batch of data (in English batch)) at the input of speech synthesizer with artificial intelligence 10. It should be noted that the dynamic linear prediction cepstral coefficients DLPCC are obtained by derivation of the LPCC coefficients using generally first- and second-order derivatives.The gain is projected via a projection tool into a space of dimension greater than its input dimension, for example of dimension equal to eight. For example, the projection tool comprises for this purpose a single-layer perceptron.The voicing is also projected via a projection tool into a space of dimension greater than its input dimension. For example, the projection tool comprises for this purpose a single-layer perceptron.Reception module 110, as illustrated by the example of figure 3, is configured to provide output 54 which comprises the outputs of the normalization and projection tools as input to projection module 12 comprising the first unidirectional recurrent neural network RN1.Moreover, reception module 110 also provides an output, comprising parameters Sv characterizing each frame of the speech synthesis suitable for being produced by an existing vocoder, as input to harmonic signal generator 14.As illustrated by the example of the architecture, projection module 12 comprises the first unidirectional recurrent neural network RN1 configured to project the parameters of output 54 grouping the elements provided by the aforementioned tools into a latent space of parameterizable dimension, for example of dimension 128 (i.e. capable of comprising 128 components).As illustrated by the example of figure 3, the first unidirectional recurrent neural network RN1 successively comprises a normalization layer LN (in English Layer-norm), followed by three dense layers, a gated recurrent unit GRU (in English Gate Recurrent Unit) and a dense layer, so as to provide as output the projection Z of said parameters in the latent space of parameterizable dimension.As a variant of the GRU unit, a layer with long short-term memory of the LSTM type (in English Long Short Term Memory) or else, more conventionally, a recurrent layer RNN (in English Recurrent Neural Network) is used.Moreover, as illustrated by the example of the architecture, harmonic signal generator 14 optionally first comprises a module 56 suitable for extracting the pitch associated with said parameters SV.Harmonic signal generator 14 further comprises a second neural network RN2 comprising in particular a Sin Gen layer, configured to generate a sinusoidal signal at the pitch associated with said received parameters, followed by a dense layer so that the signal produced by generator 14 is a signal composed of a sinusoidal signal at the pitch associated with said received parameters and its harmonics, this signal corresponding to a weighted sum E of the pitch and its harmonics. The weighting is learned by the second neural network RN2 during a prior training phase. Output E generated by harmonic signal generator 14 plays the role of the “source” part of this “source-filter” architecture.Projection model 12 and harmonic signal generator 14 are suitable for operating in parallel or successively according to any order, provided that output Z of projection model 12 and output E of harmonic signal generator 14 are provided as input and processed simultaneously by said speech synthesis module 16.From said inputs Z and E, speech synthesis module 16, via its third unidirectional recurrent neural network RN3, is configured to synthesize an encoded voice associated with said parameters SV.More precisely, speech synthesis module 16 plays the role of the “filter” part of this “source-filter” architecture of the synthesizer according to the present invention. According to the example of figure 3, the third unidirectional recurrent neural network RN3 of speech synthesis module 16 successively comprises a first dense layer corresponding in particular to a conventional perceptron and having only input, input E.This first dense layer of the third unidirectional recurrent neural network RN3 is followed by two successive filtering units F1 and F2 (in English Filter Unit). The first filtering unit F1 receives two inputs, namely both output Z of projection model 12 and the output of the first dense layer, whereas the second filtering unit F2, placed after the first filtering unit F1, also receives two inputs, namely both output Z of projection model 12 and the output of the first filtering unit F1.These two filtering units F1 and F2 have an identical structure as illustrated at zoom 57 inside said filtering unit structure.More precisely, each filtering unit F1 or F2 first comprises a concatenation layer suitable for concatenating the two received inputs represented on the one hand by the letter “y” (corresponding to output S of the first dense layer of the third unidirectional recurrent neural network RN3 for the first filtering unit F1 or to output S11 of the first filtering unit F1 for the second filtering unit F2), and on the other hand by the letter “Z” corresponding to output Z of projection model 12.At the output of said concatenation layer, each filtering unit F1 or F2 comprises a gated recurrent unit GRU (in English Gate Recurrent Unit).As a variant of the GRU unit, a layer with long short-term memory of the LSTM type (in English Long Short Term Memory) or else, more conventionally, a recurrent layer RNN (in English Recurrent Neural Network) is used.The GRU unit of each filtering unit F1 or F2 provides two outputs denoted in zoom 57 S1 and S2 at the input of each of a dense layer, the input represented by the letter “y”, as specified above, being added as illustrated within zoom 57 of Figure 3 to the output S1 of one of these two dense layers.For the first filtering unit F1, the outputs S1 and S2 indicated in zoom 57 correspond respectively to the outputs S11 and S12, whereas for the second filtering unit F2, the outputs S1 and S2 indicated in zoom 57 correspond respectively to the outputs S21 and S22.Finally, as illustrated by the example of Figure 3, the output S21 of the second filtering unit F2 is provided at the input of a last dense layer of said third unidirectional recurrent neural network RN3.Optionally, as illustrated by Figure 3, the last dense layer of said third unidirectional recurrent neural network RN3 also receives a second input equal to the sum of: the output S12 of the first filtering unit F1 to which the output S22 of the second filtering unit F2 is added, to which is furthermore added a random white noise represented by the letter a in particular to serve as a basis for the stochastic components of the voice.The output 58 of the last dense layer of said third unidirectional recurrent neural network RN3 corresponds to the encoded voice (i.e. the synthesized voice) provided by the speech synthesis module 16.Figure 4 illustrates the principle of adversarial training optionally implemented according to the present invention.More precisely, in Figure 4 the adversarial training module 60 comprises two discriminators (not represented in Figure 4) configured to be trained simultaneously with said artificial-intelligence speech synthesizer 10 during its supervised learning phase, according to a zero-sum game, said discriminators being trained to differentiate the speech syntheses provided by an existing vocoder from those 48 suitable for being synthesized by said artificial-intelligence speech synthesizer 10 of the enhancement device 100, said artificial-intelligence speech synthesizer 10 being trained to fool said two discriminators.For example, the two discriminators correspond to the discriminators D1 and D2 described in the article by J. Kong et al. entitled “HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis” NeurIPS 2020, but applied in another context without combined implementation of a speech synthesizer architecture according to a source-filter model and of the zero-sum game according to the present invention where said discriminators D1 and D2 are trained to differentiate real signals from those suitable for being synthesized by said artificial-intelligence speech synthesizer 10, said artificial-intelligence speech synthesizer 10 being trained to fool said discriminators D1 and D2.As indicated above, advantageously one of said two discriminators is sensitive to low-frequency information and the other of said two discriminators is sensitive to high-frequency information, the use of two discriminators, sensitive to low- and high-frequency information, allows the artificial-intelligence speech synthesizer 10 to correct its defects over extended frequency ranges, and is particularly effective for removing the metallic artifacts from the voice synthesized via the artificial-intelligence speech synthesizer 10.More precisely, each discriminator of the adversarial training module 60 is configured to provide a prediction PR when the incoming signal is indeed real 36 or a prediction PS when the incoming signal has been synthesized by the artificial-intelligence speech synthesizer 10 and the prediction result is then used in a loss function of the synthesizer during learning. The discriminators and synthesizers are trained jointly (i.e. joint learning).The loss function of the speech synthesizer is the average, over the different frames of the sample, of (1- PS)2 (PS being equal to one if a discriminator “says” that the signal is real and zero otherwise).The loss function of each discriminator is the sum of the averages of:(1- PR)2 and PS2 (PS and PR being equal to one if the discriminator predicts that the signal is real and zero otherwise).The loss function taking into account this prediction during the supervised learning phase of the adversarial training module, comprising the two discriminators, and jointly of the artificial-intelligence speech synthesizer 10 is advantageously adaptive. In particular, during learning, a temporal average of the spectrograms is performed, followed by a moving average over the frequencies (i.e. second axis of the spectrogram) in order to smooth the frequency distribution of the signal used for each iteration of the supervised learning.Next, this smoothed spectrum x is transformed into v by means of the following activation function: v = 1,3 - sigmoid (log10(amplitude(x))) in order to calculate frequency weighting weights, in the loss function using the spectrogram, this loss function using the spectrogram corresponding to the sum for each time / frequency value of the amplitude of the spectrogram of the L1 norm (absolute value) of the difference between the predicted spectrogram and the reference spectrogram.Such processing allows an adaptation, at each learning step, to the signal actually used and therefore allows the learning to focus on the relevant frequencies of the frame in question (i.e. the current batch of sound data (from the English batch) in order to have learning resulting in a model of better quality.A speech synthesis method 70 implemented via said artificial-intelligence speech synthesizer 10 will now be explained with reference to Figure 5 presenting a flowchart of the steps of this method 70.In general, said method 70 comprises, during its inference phase 72, a first step of receiving parameters SV produced by an existing vocoder.Optionally, this receiving step comprises a sub-step of identifying the predetermined codec used by the existing vocoder by analyzing the parameters characterizing each frame of the speech synthesis suitable for being produced by said existing vocoder, said identification sub-step further comprising the selection, from among a plurality of previously trained artificial-intelligence speech synthesizers respectively associated with distinct codecs, of the artificial-intelligence speech synthesizer corresponding to said identified predetermined codec.Said method 70 further comprises a plurality of following steps implemented by a trained artificial-intelligence speech synthesizer V_E of the enhancement device 100 having an architecture according to a neuronal source-filter model.Said plurality of following steps comprises a step 74 of projection P, via a first unidirectional recurrent neural network of said speech synthesizer, of parameters, SV, into a latent space of parameterizable dimension.Said method 70 also comprises a step 76 of generation G, via a second neural network of said trained artificial-intelligence speech synthesizer V_E, of a signal representative of the harmonics of the speech synthesis associated with said received parameters SV.The steps of projection 74 and of generation 76 of the harmonic signal are suitable for being implemented in parallel as illustrated according to Figure 5, or, not represented, successively according to any order, provided that the output of the projection step 74 and the output of the harmonic signal generation step 76 associated with one same frame of the speech synthesis associated with said received parameters SV are processed simultaneously during a subsequent step 78 of speech synthesis SYNTH as such via a third unidirectional recurrent neural network, of an encoded voice associated with said of the speech synthesis associated with said received parameters SV.Optionally, as represented by dotted lines, said inference phase 72 further comprises a step 80 of reconstruction R_T_M of the missing frames of said synthesized encoded voice from the frames preceding or following said missing frames, and / or a step 82 of band extension E_B for reconstructing a signal comprising said synthesized encoded voice with a sampling frequency higher than that of the speech synthesis associated with said received parameters SV, and / or a step 84 of extraction BIO of the vocal biomarkers of the speaker having generated said input sound signal of the existing vocoder having provided said parameters SV via at least one of said neural networks of said speech synthesizer.As an optional supplement, when the artificial-intelligence speech synthesizer (i.e. AI speech synthesizer) used has not been previously trained, the method 70 also comprises beforehand a learning phase 88 (i.e. a training phase of each element of the AI speech synthesizer).More precisely, the learning phase 88 consists in reconstructing the original communication (that is to say without alteration due to the encoding of the parameters) from the encoded (and therefore degraded) parameters. This learning mode may be enriched and allow new use cases.A new use case is obtained for example by requesting (i.e. by training to) reconstruct, via the aforementioned optional reconstruction module 20, the frames lost during transmission from a context (i.e. the frames preceding and following the lost frames), which makes it possible to reconstruct part of the lost communication.Another new use case is obtained by requesting (i.e. by training to) reconstruct an audio signal with a sampling rate higher than that of the original codec via the aforementioned optional band extension module 22, which makes it possible to improve the perception of voice quality and robustness to low-frequency noises. For example, MELPe 1200 is designed to reconstruct audio signals sampled at 8kHz and the AI speech synthesizer of the speech synthesis improvement device 100 is trained to reconstruct signals sampled at 16kHz.According to an optional aspect of this optional complement, said learning phase 88 comprises a step 90 of adversarial training E_A via said previously described adversarial training module.A person skilled in the art will understand that the invention is not limited to the embodiments described, nor to the particular examples of the description, the embodiments and variants mentioned above being suitable for being combined with one another in order to generate new embodiments of the invention.The present invention thus makes it possible to “replace” the speech synthesis part of an existing vocoder, and to improve the rendering of the voice communication corresponding to the received transmissionFrom a hardware point of view, the proposed solution is suitable for taking place in all of the existing equipment used a codec (e.g. portable military or civilian radio, on a fixed station, etc.)Moreover, optionally, as mentioned above, the present invention is suitable for offering new functionalities compared with existing vocoders by allowing, via the reconstruction module 20, a reconstruction (from the English inpainting) of the missing frames of the speech synthesis which would have been produced by an existing vocoder in order in particular to fill micro-cuts, and / or further, via the band extension module 22, a band extension without changing the codec of the already existing vocoders, which allows interoperability with the radio stations already transmitting in the identified codec.An improvement in the quality of speech synthesis at constant bit rate is thus obtained and the solution according to the present invention is suitable for reaching, or even exceeding, the performance of vocoders twice as greedy in bit rate as the existing vocoder having provided it at the input with the parameters characterizing each frame of the speech synthesis suitable for being produced by said existing vocoder. The reduction in bit rate is also accompanied by new performance gains related to the communication. Indeed, at a lower bit rate, the range of the transmitting stations is increased. A lower bit rate makes it possible, at constant transmission capacity, to allocate space to other transmissions.The present invention moreover makes it possible, as experimentally verified, by virtue of the specific structure of the artificial intelligence speech synthesizer of the improvement device according to the present invention, to provide an improvement compared with current vocoding with equivalent input parameters.In particular, the source-filter-type neural architecture of the speech synthesizer according to the present invention makes it possible to have a “light” solution (i.e. frugal, not very greedy) in computational and memory footprint compared with current AI solutions, integrable within embedded products, and opens up the possibility of implementing other “functionalities” such as, for example, reconstruction by Inpainting, band extension (from the English bandwidth extensionbandwidth extension[[[ / DOC_STYLE:2]]]) (this involves an addition of information by prediction) or also “voice conversion” (i.e. passage from one voice to another voice (a priori associated with existing speakers) or also “voice transformation) (i.e. passage from one voice to another voice (which may be, more generally, that of a virtual speaker)).By delegating speech synthesis to AI approaches with recurrent neural networks at each stage of the source-filter model rather than to solutions designed by “experts” (i.e. solutions resulting from modelling), the solution according to the present invention makes it possible in particular to compensate for a potential loss of information related to the reduction in the number of input parameters, in particular in the context of a low-bit-rate application.The choice of a “biased” source-filter architecture (i.e. an architecture of neural networks which has been “biased” or rather specialized towards speech synthesis by adding a modelling of speech by a sum of harmonics) towards speech synthesis makes it possible to reduce the digital footprint, in terms of memory and computation, of the implemented neural networks.Finally, optionally, the practical difficulty of training neural networks in such an open task with few input parameters for a large number of possible temporal signals is advantageously mitigated by the use of adversarial training in which two discriminators are trained to recognize the signals generated by the speech synthesizer according to the present invention or jointly said speech synthesizer is trained to deceive the discriminators.Thus, the device and the method for improving speech synthesis are suitable for being integrated within the framework of digital radiocommunication applications, voice over IP (from the English Internet Protocol) VoIP, audio compression, etc.
Claims
1. Device (100) for improving the speech synthesis produced, by means of a predetermined codec, by an existing vocoder, said device comprising:- a module (110) for receiving the parameters characterizing each frame of the speech synthesis suitable for being produced by said existing vocoder;- at least one artificial intelligence speech synthesizer (10) having an architecture according to a neural source-filter model, said artificial intelligence speech synthesizer comprising at least:- a projection module (12) configured to project, via a first unidirectional recurrent neural network, said parameters, or parameters derived from said parameters, into a latent space of configurable dimension;- a harmonic signal generator (14) configured to generate, via a second neural network, a signal representative of the harmonics of said speech synthesis associated with said parameters received via said reception module (110);- a speech synthesis module (16) configured to synthesize, via a third unidirectional recurrent neural network, an encoded voice associated with said human voice from:- the projection of said projected parameters into said latent space of configurable dimension provided by said projection module, and- said signal representative of the harmonics of the speech synthesis associated with said parameters received via said reception module (110) and provided by said generator;said artificial intelligence speech synthesizer (10) having been previously trained during a supervised learning phase using training speech syntheses produced by means of said predetermined codec.
2. Device (100) according to claim 1, wherein the reception module (110) comprises an identification tool for said predetermined codec by analyzing the speech synthesis produced by said existing vocoder, said identification tool being further configured to select, from among a plurality of artificial intelligence speech synthesizers previously trained and respectively associated with distinct codecs, said artificial intelligence speech synthesizer (10) corresponding to said identified predetermined codec.
3. Device (100) according to claim 1 or 2, wherein the reception module (110) comprises an interpolation tool configured to linearly interpolate said parameters and provide as input to said artificial intelligence speech synthesizer (10) interpolated parameters compatible with said speech synthesis module (16).
4. Device (100) according to any one of the preceding claims, wherein said signal generated by the generator (14) and representative of the harmonics of said speech synthesis associated with said parameters received via said reception module (110) is a weighted sum of the signal corresponding to the pitch associated with said received parameters and its harmonics, said weighting having been learned beforehand by said second neural network.
5. Device (100) according to claim 3, wherein white noise is added to said weighted sum.
6. Device (100) according to any one of the preceding claims, wherein the artificial intelligence speech synthesizer (10) further comprises an adversarial training module (18) comprising two discriminators configured to be trained simultaneously with said artificial intelligence speech synthesizer during its supervised learning phase, according to a zero-sum game, said discriminators being trained to differentiate real signals from those suitable for being synthesized by said artificial intelligence speech synthesizer, said artificial intelligence speech synthesizer being trained to deceive said two discriminators.
7. Device (100) according to claim 5, wherein one of said two discriminators is sensitive to low-frequency information and the other of said two discriminators is sensitive to high-frequency information.
8. Device (100) according to any one of the preceding claims, wherein the artificial intelligence speech synthesizer (10) further comprises:- a reconstruction module (20) configured to reconstruct missing frames of said synthesized encoded voice from the frames preceding or following said missing frames, and / or- a band extension module (22) configured to reconstruct a signal comprising said synthesized encoded voice with a sampling frequency higher than that of said speech synthesis associated with said parameters received via said reception module (110).
9. Method (70) for improving the speech synthesis produced, by means of a predetermined codec, by an existing vocoder, said method being implemented by an improvement device (100) according to any one of the preceding claims, said method comprising:- receiving the parameters characterizing each frame of the speech synthesis capable of being produced by said existing vocoder;and the following steps implemented during an inference phase by an artificial-intelligence speech synthesizer (10) of said improvement device (100), said artificial-intelligence speech synthesizer (10) having been previously trained during a supervised learning phase (88) using training parameters characterizing each frame of speech syntheses capable of being produced by means of said predetermined codec:- projection (74), via a first unidirectional recurrent neural network of said artificial-intelligence speech synthesizer, of the parameters received via said reception module (110) or of parameters derived from said parameters, into a latent space of configurable dimension;- generation (76), via a second neural network of said artificial-intelligence speech synthesizer (10), of a signal representative of the harmonics of the speech synthesis associated with said parameters received via said reception module (110);- speech synthesis (78), via a third unidirectional recurrent neural network, of an encoded voice associated with said human voice from:- the projection of said projected parameters into said latent space of configurable dimension provided by said projection module, and- said signal representative of the harmonics of the speech synthesis associated with said parameters received via said reception module (110) and provided by the generator of the artificial-intelligence speech synthesizer (10) of said improvement device (100).
10. Program comprising software instructions which, when executed by a computer, implement a method for improving speech synthesis according to claim 9.