Method and device for estimating the pitch of a human voice in an audio signal

Abstract

The present invention relates to a method (50) for estimating the pitch of a human voice of an audio signal received as input, said method being implemented by an electronic device for estimating the pitch of a human voice and comprising the following steps:- decomposing (52) said signal into frames of predetermined finite size;- for each frame,- calculating (54):- an autocorrelation of said frame; and- a spectral transformation of said frame;- providing (62) the pair of results of said calculation as input to a neural network with an attention mechanism, previously trained, and configured to provide at the output the probability that the autocorrelation lag corresponds to the pitch period of said voice of the audio signal received as input;- from said probability, estimating (76) said pitch.

Description

 TITLE: Method and device for estimating the pitch of a human voice in an audio signal The present invention relates to a method for estimating the pitch of a human voice of an audio signal received as input, said method being implemented by an electronic device for estimating the pitch of a human voice.The invention also relates to a computer program comprising software instructions which, when implemented by a programmable electronic device, implement such a method.The invention also relates to an electronic device for estimating the pitch of a human voice of an audio signal received as input. The invention lies in the field of analysis of the human voice and / or vocoding (i.e. the synthesis or alternatively the coding of the human voice), these fields often using the voice information called pitch or tone height corresponding, according to the French Association for Standardization AFNOR, to the "character of the auditory sensation related to the frequency of a periodic sound, which makes the sound high-pitched or low-pitched depending on whether this frequency is higher or lower", and often considered to be equivalent to the fundamental frequency of the voice.In the technical fields of the invention, the pitch makes it possible to analyze the voice, in order to identify bio-information and to allow, for example, speaker recognition. In the field of vocoding, that is to say the coding of the voice to facilitate voice communications (i.e. radiocommunication, VoIP, etc.), the correct estimation of the pitch allows better quality of voice restitution.However, the pitch is not strictly speaking equivalent to the fundamental frequency of the voice because, as mentioned above, it is a subjective psychoacoustic property which may depend on the frequency context of the sound.The uncertainty associated with the estimation of the pitch decreases as the listening interval (i.e. the input frame) increases. This decrease is physical and is related to the period / frequency duality of the sound signal comprising the recording of the human voice. However, the decrease in the listening interval is beneficial for applications related to the human voice.The acoustic noises surrounding the recording of the human voice are moreover detrimental to the precise estimation of its pitch. Indeed, the acoustic noises may alter the frequency domain of the pitch to be estimated.As a result, the estimation of the pitch is a difficult task, due to the subjectivity mentioned above, and due to the fundamental frequency being easily liable to be confused with one of its harmonics.The current conventional pitch estimation techniques are moreover not very robust to environmental acoustic noises. Indeed, such conventional methods mitigate the problem by making assumptions about the environmental noise or noises, such as the type of noise, for example Gaussian, or alternatively the frequency range over which the noise extends, and these assumptions do not correspond to reality.Moreover, conventional methods are based on a mathematical shortcut of the pitch, by reducing it to the fundamental frequency of the voice. Although this definition is justified in a majority of cases, it does not make it possible to estimate the pitch under certain difficult conditions, because the pitch may be related to the context of the sound (i.e. to its environment).To remedy this, artificial intelligence techniques, in particular via machine learning and especially deep learning (from the English deep learning) have been proposed.However, such techniques are demanding in terms of numerical complexity, because they are generally based on neural networks using a large number of internal parameters and numerous mathematical operations, which makes them unsuitable for use in a hardware environment constrained in memory and / or in power.The aim of the invention is therefore to propose a method for estimating the pitch of the human voice which is precise, robust to acoustic noises and of low numerical complexity.For this purpose, the invention relates to a method for estimating the pitch of a human voice of an audio signal received as input, said method being implemented by an electronic device for estimating the pitch of a human voice and comprising the following steps:- decomposing said signal into frames of predetermined finite size;- for each frame,- calculating:- an autocorrelation of said frame; and- a spectral transformation of said frame;- supplying the pair of results of said calculation as input to a neural network with an attention mechanism, previously trained, and configured to supply as output the probability that the shift of the autocorrelation corresponds to the period of the pitch of said voice of the audio signal received as input;- from said probability, estimating said pitch.Unlike the aforementioned prior art according to which conventional methods focus solely on one piece of information, and according to which artificial intelligence techniques which give the neural networks the choice of the basis on which the input information is projected, in particular during their learning phase, the present invention proposes using temporal and frequency information simultaneously.Such simultaneous (i.e. joint) use of the temporal and frequency information of the frame makes it possible to obtain, by crossing them, a precise and noise-robust estimation of the pitch of the human voice while being associated with low complexity.According to other advantageous aspects of the invention, the method for estimating the pitch of the human voice comprises one or more of the following features, taken in isolation or according to all technically possible combinations:- said spectral transformation is a spectral transformation belonging to the group comprising at least:- a constant-Q transform;- a discrete cosine transform suitable for providing Mel-frequency cepstral coefficients;- a transform based on the amplitude of a fast Fourier transform;- the method comprises a preliminary step of partitioning the latent space of said neural network;- said partitioning is a k-means partitioning or a partitioning via Gaussian mixture models;- the autocorrelation lag corresponding to the maximum of said probability is selected and then used with the sampling frequency of said audio signal to determine the fundamental frequency associated with the pitch of said voice, said fundamental frequency corresponding to the ratio between said sampling frequency and said selected lag;- the method further comprises a step of correcting said probability by implementing a Viterbi algorithm applied at the output of said neural network and suitable for taking into account the evolution of said pitch over several successive frames;- said neural network with an attention mechanism is a neural network of the “pointer network” type and is previously trained by means of supervised training;- said calculation of said spectral transformation further comprises a formatting step, by means of a convolutional neural network, suitable for making the result of said calculation of said spectral transformation concatenable with the result of said calculation of said autocorrelation of said frame to form, via concatenation during said provision, said pair of results provided as input to said neural network with an attention mechanism.The invention also relates to a computer program comprising software instructions which, when executed by a computer, implement a method for estimating the pitch of the human voice as defined above.The invention also relates to an electronic device for estimating the pitch of a human voice from an audio signal received as input, said electronic device being characterized in that it comprises:- a decomposition module configured to decompose said signal into frames of predetermined finite size;- a calculation module configured to calculate, for each frame:- an autocorrelation of said frame; and- a spectral transformation of said frame;- a module for providing the pair of results of said calculation as input to a neural network with an attention mechanism, previously trained, and configured to provide the probability that the autocorrelation lag corresponds to the pitch period of said voice of the audio signal received as input;- said neural network with an attention mechanism;- an estimation module configured to estimate said pitch from said probability provided at the output by said neural network with an attention mechanism.The invention will become clearer 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 schematic view of an electronic device for estimating the pitch of the human voice according to the present invention;- [Fig. 2] figure 2 is a schematic view of an example of an architecture of the device for estimating the pitch of the human voice according to the present invention;- [Fig. 3] figure 3 is a flowchart of a method for estimating the pitch of the human voice according to an embodiment of the present invention. The electronic device for estimating the pitch of the human voice from an audio signal received as input according to the present invention is illustrated in figure 1.As is visible in the embodiment of figure 1, the electronic device for estimating the pitch of the human voice from an audio signal received as input comprises a decomposition module 12 configured to decompose said signal into frames of predetermined finite size connected at the output to the input an calculation module 14 configured to calculate, for each frame, an autocorrelation of said frame and a spectral transformation of said frame.Said calculation module 14 is connected at the output to the input of a provision module 16, of the pair of results of said calculation, as input to a neural network 18 with an attention mechanism, previously trained, and configured to provide at the output the probability that the autocorrelation lag corresponds to the pitch period of said voice of the audio signal received as input.Said neural network with an attention mechanism 18 is connected at the output to the input of an estimation module 20 configured to estimate said pitch from said probability provided at the output by said neural network with an attention mechanism 18 as represented in figure 1.As an optional addition, the electronic device for estimating the pitch of the human voice from an audio signal received as input comprises a correction module 22, the input of which is connected to the output of said neural network with an attention mechanism 18. Such an optional correction module 22 is configured to correct said probability by implementing a Viterbi algorithm applied at the output of said neural network and suitable for taking into account the evolution of said pitch over several successive frames.In the example of figure 1, the electronic device for estimating the pitch of the human voice from an audio signal received as input comprises an information processing unit 24 formed, for example, by a memory 26 and a processor 28 associated with the memory 26.In the example of figure 1, the decomposition module, the calculation module and the provision module, the neural network and the estimation module, as well as, as an optional addition, the correction module, are each produced in the form of software, or of a software block, executable by the processor. The memory of the electronic device for estimating the pitch of the human voice from an audio signal received as input is then suitable for storing decomposition software, calculation software, provision software, software for implementing a neural network and estimation software, as well as, as an optional addition, correction software. The processor is then suitable for executing each of the software items among the decomposition software, the calculation software, the provision software, the software for implementing a neural network and the estimation software, as well as, as an optional addition, the correction software.In an alternative embodiment not shown, the decomposition module, the calculation module, the provision module, the neural network and the estimation module, as well as, as an optional addition, the correction module, are each produced in the form of a programmable logic component, such as an FPGA (of the English Field Programmable Gate Array), an integrated circuit, such as an ASIC (of the English Application Specific Integrated Circuit), one or more GPUs (of the English Graphics Processing Unit), or one or more accelerators dedicated to artificial intelligence AI.When the electronic device for estimating the pitch of the human voice from an audio signal received as input 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 furthermore suitable for being recorded on a non-illustrated, computer-readable medium. 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 disk, a magneto-optical disk, 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.Some of these modules and, where applicable, their implementation options are described in greater detail below.More precisely, the decomposition module 12 is suitable for decomposing the input audio signal into frames of predetermined finite size, for example of 22.5 ms, 67.5 ms (i.e. 3x22.5 ms), 90 ms, 20 ms, 30 ms, etc.The calculation module 14 subjects each frame of finite size to two independent transformations.Indeed, the calculation module 14 is, as indicated above, suitable for calculating, for each frame of the audio signal provided by the decomposition module 12, on the one hand an autocorrelation A of said frame; and on the other hand a spectral transformation S of said frame.As an optional addition, said spectral transformation is a spectral transformation belonging to the group comprising at least:- a constant-Q transform (CQT);Constant-Q transform- a discrete cosine transform suitable for providing Mel-frequency cepstral coefficients MFCC (from the English Mel-frequency cepstral coefficients).- an amplitude-based transform of a fast Fourier transform FFT (from the English Fast Fourier Transform);- etc.Preferably, the constant-Q transform is used because it has been experimentally shown to be more effective than the amplitude-based transform of a fast Fourier transform.It should be noted that the spectral transformation of the frame is defined before the training phase of the neural network 18.The provision module 16 is configured to simultaneously provide the pair of results delivered by the computation module 14 as input to the neural network 18, said pair comprising the autocorrelation result A of said frame and the result S of said spectral transformation.In particular, to form said pair, the provision module 16 is suitable for concatenating, within one and the same input of the neural network 18, the autocorrelation result of said frame and the result of said spectral transformation. It should be noted that the order of concatenation implemented during the learning is arbitrary, the autocorrelation result of said frame being capable of being before or after the result of said spectral transformation, but the order of concatenation implemented during the learning freezes that implemented subsequently during inference.Via said pair at the input, the results A and S of the two transformations are then processed simultaneously by the neural network 18 which, based on a self-attention mechanism, will provide the probability that the offset (in English lag) of the autocorrelation A corresponds to the Pitch period of the voice.The probability is a conditional probability given both the results S and A of the two transformations, respectively of the spectral transformation and of the autocorrelation of the frame.Subsequently, it is considered that the neural network 18 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 can be envisaged with a layer that can be connected to a layer more distant than the immediately preceding layer.Each neuron is also associated with an operation, that is to say a type of processing, to be carried out 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 applying an activation function, typically a non-linear function, to said weighted sum, and 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 additionally 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 designation « dense layer ».As an optional addition, the neural network 18 is based on a neural network architecture used in the field of automatic language processing, the « Pointer Networks » such as notably introduced by O. Vinyals et al. in the article « Pointer Networks » of June 9, 2015.The use of a « Pointer Networks »-type architecture in the present field of pitch detection is advantageous because it makes it possible to achieve gains in memory due to a limited number of internal parameters, in numerical complexity, with a limited number of mathematical operations.Experimentally, such gains have made possible the implementation of the solution according to the present invention in an extremely constrained hardware environment with an electrical consumption of less than 0.05W.A training phase of the neural network 18, before implementation of the solution, makes it possible to make the solution robust to acoustic noise. More precisely, said neural network, preferably of the « pointer networks » type, is previously trained via a supervised training phase, using data annotated by a human expert, which makes it possible, in particular, to no longer restrict the pitch to the fundamental frequency of the human voice.As an optional addition, the latent space of said neural network is suitable for being, beforehand, partitioned according to a partitioning (also called clustering from the English data clustering).According to a first variant of this optional addition, said partitioning is a partitioning via an unsupervised method such as a k-means partitioning (or k-means in English) or a partitioning via Gaussian Mixture Models GMM (from the English Gaussian Mixture Models).Such prior partitioning of the latent space of the neural network 18 makes it possible to detect transitions of voiced frames (i.e. comprising human voice) from those that are unvoiced (i.e. comprising no human voice).More precisely, the principle is to force a partition into two (after having empirically verified it on the data) and each partition corresponds to voiced or unvoiced.More precisely, the determination of the unvoiced / voiced transition is said to be semi-supervised because obtained both with the aid from a partitioning (clustering) according to an unsupervised method and with the aid of a supervised training phase of said neural network 18.The present invention thus proposes to take advantage of the use of the latent space of the neural network 18, its partitioning thus advantageously making it possible to provide a voiced / unvoiced classification of the frame.According to a first embodiment, the output of the neural network 18 with an attention mechanism is directly connected to the input of an estimation module 20 configured to estimate said pitch from said probability provided at the output by said neural network with an attention mechanism 18.Within the scope of this first embodiment (i.e. without correction processing / processing before estimation), the estimation module 20 is suitable for selecting the autocorrelation offset corresponding to the maximum of said probability and then using it, with the sampling frequency of said audio signal, to at least determine the fundamental frequency associated with the pitch of said voice, said fundamental frequency corresponding to the ratio between said sampling frequency and said selected offset.Optionally, according to a second embodiment, the optional correction module 22 is implemented between the output of the neural network 18 and the input of the estimation module 20. The optional correction module 22 is configured to correct said probability by implementing a Viterbi algorithm applied at the output of said neural network 18 and is suitable for taking into account the evolution of said pitch over several successive frames.Indeed, the use according to the present invention of a probabilistic model at the output of the neural network 18 allows the use, by the optional correction module 22, of a Viterbi-type algorithm to make the estimation of the pitch more robust by taking into account the « trajectory » (i.e. the temporal evolution) of the pitch over several successive frames.Figure 2 is a schematic view of a detailed example of architecture of the device 10 for estimating the pitch of the human voice according to the present invention.The modules identical to those in Figure 1 are shown in Figure 2 with identical numbering.According to this example of architecture of the device 10 for estimating the pitch of the human voice, the decomposition module 12 receives as input the input audio signal E which it is capable of decomposing into frames TR of predetermined finite size.The calculation module 14 comprises, as illustrated by Figure 2, two calculation tools 30 and 31 implemented in parallel according to the example of Figure 2.The first calculation tool 30 is configured to calculate the autocorrelation A of each frame supplied by the decomposition module 12.The second calculation tool 31 is configured to apply a spectral transformation to each frame supplied by the decomposition module 12. More precisely, according to the example of Figure 2, the spectral transformation 32 implemented by the calculation tool 31 and defined before training of the neural network 18 corresponds to a constant-Q transform CQT.As an optional addition, as illustrated by the example of Figure 2, the calculation tool 31 of the spectral transformation further comprises a previously trained convolutional neural network 34, capable of advantageously implementing a formatting F, capable of making the result S supplied by said calculation tool 31 of said spectral transformation concatenable with the result A supplied by said autocorrelation calculation tool 30 of said frame.A convolutional neural network is also sometimes called a convolution neural network or by the acronym CNN, which refers to the English designation of “Convolutional Neural Networks”.In a convolutional neural network, each neuron of the same layer presents exactly the same connection pattern as its neighboring neurons, but at different input positions. The connection pattern is called a convolution kernel or, more often, “kernel” with reference to the corresponding English designation.More precisely, said convolutional neural network 34 comprises, according to the example of architecture of Figure 2, a dense layer, NC convolutional layers called ConvB, with NC an integer, a formatting layer F (flatten in English) and a dense layer.According to other examples not shown, the two calculation tools 30 and 31 are implemented successively, the autocorrelation calculation tool 30 being implemented before the spectral transformation calculation tool 31, or vice versa. It should be noted that once this order is fixed during learning, the same order must be used during inference.The results A and S of the calculation module 14 are supplied respectively, for the same frame, by the two independent calculation tools 30 and 31 at the input of the supply module 16 configured to simultaneously supply, via concatenation, during said supply, the pair of results delivered by the calculation module 14 at the input of the neural network 18 with attention mechanism.More precisely, according to the example of Figure 2, the neural network 18 is preferentially of the “Pointer Networks” type and has two distinct stages 36 and 38. The first stage 36 corresponds to an encoder stage composed in particular of at least one dense layer and at least one gated recurrent unit GRU (from the English Gate Recurrent Unit).The second stage 38 is dedicated to the attention mechanism corresponding in particular to a self-attention mechanism and composed of at least two dense layers capable of translating the similarity between the autocorrelation A lag and the pitch period, while knowing the frequency representation S of the frame.At the output of this self-attention stage 38, the similarity is transformed via a probabilistic model, in particular by means of the mathematical function softmax 40, into the probability 42 that the lag (in English lag) of the autocorrelation A corresponds to the Pitch period of the voice.This probability 42 is then used, according to the example of Figure 2, directly by the pitch estimation module 20.In other words, the example of Figure 2 illustrates the aforementioned first embodiment, where the output of the neural network 18 with attention mechanism is directly connected to the input of an estimation module 20 configured to estimate said pitch from said probability supplied at the output by said neural network with attention mechanism 18.Within the framework of this first embodiment (i.e. without correction processing / processing before estimation), the estimation module 20 is capable of selecting the autocorrelation lag corresponding to the maximum of said probability and then using it, with the sampling frequency of said audio signal, to at least determine the fundamental frequency associated with the pitch of said voice, said fundamental frequency corresponding to the ratio between said sampling frequency and said selected lag.An example of an embodiment of the operation, according to the present invention, of such a device 10 is described below in relation to Figure 3.More precisely, the method 50 for estimating the pitch of a human voice of an audio signal received as input first comprises a first step 52 of decomposing said signal into frames TR of predetermined finite size.Then, the method 50 comprises a step 54 of calculating two independent transformations of each frame resulting from the decomposition step 52.More precisely, said calculation step 54 comprises two independent calculation sub-steps 56 and 58, namely a sub-step 56 of calculating C_A an autocorrelation of said frame, and a sub-step 58 of calculating C_T_S a spectral transformation of said frame.As an optional addition, said spectral transformation 68 is a spectral transformation belonging to the group comprising at least:- a constant-Q transform;- a discrete cosine transform capable of providing Mel-frequency cepstral coefficients;- a transform based on the amplitude of a fast Fourier transform.These two sub-steps 56 and 58 are capable of being implemented in parallel, as illustrated by Figure 3, or successively in any order.According to an optional aspect, as indicated above, said spectral transformation 58 further comprises a formatting step 60, by means of a convolutional neural network, capable of making the result of said calculation of said spectral transformation concatenable with the result of said autocorrelation calculation of said frame in order to form, via concatenation during said supply, said pair of results supplied at the input of said neural network with attention mechanism.Then the method 50 comprises a step 62 of supplying the pair of results of said calculation at the input of a step 64 of inference phase P_I of the neural network with attention mechanism 18, said pair being formed in particular via concatenation.As illustrated by Figure 3, the neural network with attention mechanism 18 used during the inference phase 64 is configured to supply at the output the probability 42 that the autocorrelation lag corresponds to the pitch period of said voice of the audio signal received as input, and was obtained following a prior step 66 corresponding to a training phase P_E of said neural network with attention mechanism.As an optional addition, prior to said training 66, the method comprises a step 68 of partitioning (clustering C) the latent space of said neural network, said partitioning 68 being a k-means partitioning or a partitioning via Gaussian Mixture Models GMM (from the English Gaussian Mixture Models).Preferentially, the aforementioned second embodiment is implemented according to the dotted arrow 70 at the output of the inference phase of the neural network with attention mechanism 18, with an optional step 72 of correcting said probability by implementation 74 of a Viterbi algorithm applied at the output of said neural network and capable of taking into account the evolution of said pitch over several successive frames.The corrected probability resulting from this optional correction step 72 is then used during a step 76 of estimating said pitch.As an alternative according to arrow 78, according to the aforementioned first embodiment, the probability 42 resulting from the inference phase 64 is directly used ((i.e. without subsequent processing for correction of the Viterbi algorithm type) during step 76 of estimating said pitch.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 the variants mentioned above being suitable for being combined with one another to generate new embodiments of the invention.The present invention thus allows use in low-bit-rate vocoders, for example due to the cross-use of autocorrelation information and spectral information (any, but defined when defining the neural network), in other words due to the use of two types of information on the frame.Moreover, the use of a specific neural network architecture of the “Pointer Networks” type makes it possible to have very small neural networks compared with the state of the art, which allows low complexity and a fast inference phase that consumes few resources.Furthermore, as indicated, the learning phase (i.e. training) of the neural network is suitable for being adapted, via semi-supervised training, to operate in a noisy environment, which makes it possible to go beyond the mathematical definition of the Pitch reduced to the fundamental frequency of the voice.The solution proposed according to the present invention is thus robust to noise owing to its training phase because the implementation of strong assumptions regarding the properties of the noise is avoided. 

Claims

1. Method (50) for estimating the pitch of a human voice from an audio signal received at input, said method being implemented by an electronic device for estimating the pitch of a human voice and comprising the following steps:- decomposing (52) said signal into frames of predetermined finite size;- for each frame,- calculating (54):- an autocorrelation of said frame; and- a spectral transform of said frame;- providing (62) the pair of results of said calculation as input to a neural network with an attention mechanism, previously trained, and configured to provide at output the probability that the autocorrelation lag corresponds to the pitch period of said voice of the audio signal received at input;- from said probability, estimating (76) said pitch,the autocorrelation lag corresponding to the maximum of said probability being selected and then used (78) with the sampling frequency of said audio signal to determine the fundamental frequency associated with the pitch of said voice, said fundamental frequency corresponding to the ratio between said sampling frequency and said selected lag.

2. Method (50) for estimating according to claim 1, wherein said spectral transform is a spectral transform belonging to the group comprising at least:- a constant-Q transform;- a discrete cosine transform suitable for providing Mel-frequency cepstral coefficients;- a transform based on the amplitude of a fast Fourier transform.

3. Estimation method (50) according to claim 1 or 2, comprising a prior step of voiced or unvoiced partitioning (68) of the latent space of said neural network.

4. Estimation method (50) according to claim 3, wherein said partitioning (68) is a k-means partitioning or a partitioning via Gaussian mixture models.

5. Estimation method (50) according to any one of the preceding claims 1 to 4, further comprising a step of correcting (72) said probability by implementing a Viterbi algorithm (74) applied at the output of said neural network and suitable for taking into account the evolution of said pitch over several successive frames.

6. Estimation method (50) according to any one of the preceding claims, wherein said neural network with an attention mechanism is a neural network of the “pointer network” type and is previously trained by means of supervised training.

7. Estimation method (50) according to any one of the preceding claims, wherein said calculation of said spectral transform further comprises a step (60) of formatting, by means of a convolutional neural network, the result of said calculation of said spectral transform.

8. Computer program comprising software instructions which, when executed by a computer, implement a method for estimating the pitch of a voice from an audio signal received at input according to any one of the preceding claims.

9. Electronic device (10) for estimating the pitch of a human voice from an audio signal received at input, said electronic device being characterized in that it comprises:- a decomposition module (12) configured to decompose said signal into frames of predetermined finite size;- a calculation module (14) configured to calculate, for each frame:- an autocorrelation of said frame; and- a spectral transform of said frame;- a module (16) for providing the pair of results of said calculation as input to a neural network with an attention mechanism, previously trained, and configured to provide the probability that the autocorrelation lag corresponds to the pitch period of said voice of the audio signal received as input;- said neural network with an attention mechanism;- an estimation module (20) configured to estimate said pitch from said probability provided at the output by said neural network with an attention mechanism,the autocorrelation lag corresponding to the maximum of said probability being selected and then used with the sampling frequency of said audio signal to determine the fundamental frequency associated with the pitch of said voice, said fundamental frequency corresponding to the ratio between said sampling frequency and said selected lag.

Images