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

Abstract

Method and device for estimating the pitch of a human voice of an audio signal 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: - decomposition (52) of said signal into frames of predetermined finite size; - for each frame, - calculation (54): - of an autocorrelation of said frame; and - of a spectral transformation of said frame; - supplying (62) the pair of results of said calculation as input to a neural network with 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, estimation (76) of said pitch. Figure for abstract: Figure 3

Description

Title of the invention: Method and device for estimating the pitch of a human voice from an audio signal

[0001] The present invention relates to a method for estimating the pitch of a human voice from an audio signal received as input, said method being implemented by an electronic device for estimating the pitch of a human voice.

[0002] The invention also relates to a computer program comprising software instructions which, when implemented by a programmable electronic device, implement such a method.

[0003] The invention also relates to an electronic device for estimating the pitch of a human voice from an audio signal received at input.

[0004] The invention lies in the field of human voice analysis and / or vocoding (i.e. synthesis or coding of human voice), these fields often using voice information called pitch or tonal height corresponding, according to the French Standardization Association AFNOR, to the "character of the auditory sensation linked 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 equivalent to the fundamental frequency of the voice.

[0005] In the technical fields of the invention, the pitch makes it possible to analyze the voice, to identify bio-information and allow, for example, the recognition of the speaker. In the field of vocoding, that is to say the coding of the voice to facilitate voice communications (i.e. radio communication, VoIP, etc.), the good estimation of the pitch allows a better quality of restitution of the voice.

[0006] 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 can depend on the frequency context of the sound.

[0007] The uncertainty related to the pitch estimation 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.

[0008] The acoustic noise surrounding the recording of the human voice is also detrimental to the precise estimation of its pitch. Indeed, acoustic noise can alter the frequency domain of the pitch to be estimated.

[0009] Therefore, estimating the pitch is a difficult task, due to the subjectivity mentioned above, the fundamental frequency being easily confused with one of its harmonics.

[0010] Current conventional pitch estimation techniques are also not very robust to environmental acoustic noise. Indeed, such conventional methods mitigate the problem by making assumptions about the environmental noise(s) such as the type of noise, for example Gaussian, or the frequency range over which the noise is spread, and these assumptions do not correspond to reality.

[0011] Furthermore, classical methods are based on a mathematical shortcut of the pitch, by reducing it to the fundamental frequency of the voice. If this definition is justified in a majority of cases, it does not allow the pitch to be estimated in certain difficult conditions, because the pitch can be linked to the context of the sound (i.e. to its environment).

[0012] To remedy this, artificial intelligence techniques, notably via machine learning and in particular deep learning, have been proposed.

[0013] 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 power.

[0014] The aim of the invention is then to propose a method for estimating the pitch of the human voice which is precise, robust to acoustic noise and of low numerical complexity.

[0015] To this end, the invention relates to a method for estimating the pitch of a human voice of 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:

[0016] - decomposition of said signal into frames of predetermined finite size;

[0017] - for each frame,

[0018] - calculation:

[0019] - of an autocorrelation of said frame; and

[0020] - a spectral transformation of said frame;

[0021] - providing the pair of results of said calculation as input to a neural network with attention mechanism, previously trained, and configured to provide as output the probability that the autocorrelation shift corresponds to the period of the pitch of said voice of the audio signal received as input;

[0022] - from said probability, estimation of said pitch.

[0023] Unlike the aforementioned state of the art according to which the conventional methods focus only on one piece of information, and according to which the 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 to use the temporal and frequency information simultaneously.

[0024] 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.

[0025] 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 characteristics, taken individually or in all technically possible combinations:

[0026] - said spectral transformation is a spectral transformation belonging to the group comprising at least:

[0027] - a transform with constant Q;

[0028] - a discrete cosine transform suitable for providing cepstral coefficients at Mel frequency;

[0029] - a transform based on the amplitude of a fast Fourier transform;

[0030] - the method comprises a prior step of partitioning the latent space of said neural network;

[0031] - said partitioning is a k-means partitioning or a partitioning via Gaussian mixture models;

[0032] - the shift of the autocorrelation 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 offset;

[0033] - the method further comprises a step of correcting said probability by betting implementing a Viterbi algorithm applied to the output of said neural network and capable of taking into account the evolution of said pitch over several successive frames;

[0034] - said neural network with attention mechanism is a neural network of type “pointer network” and is previously trained using supervised training;

[0035] - said calculation of said spectral transformation further comprises a step of formatting, 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 to form, via concatenation during said providing said pair of results provided as input to said neural network with attention mechanism.

[0036] 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.

[0037] The invention also relates to an electronic device for estimating the pitch of a human voice from an audio signal received at input, said electronic device being characterized in that it comprises:

[0038] - a decomposition module configured to decompose said signal into frames of predetermined finished size;

[0039] - a calculation module configured to calculate, for each frame:

[0040] - an autocorrelation of said frame; and

[0041] - a spectral transformation of said frame;

[0042] - a module for providing the pair of results of said calculation as input to a network of neurons with attention mechanism, previously trained, and configured to provide the probability that the autocorrelation shift corresponds to the period of the pitch of said voice of the audio signal received as input;

[0043] - said neural network with attention mechanism;

[0044] - an estimation module configured to estimate said pitch from said probability provided as output by said neural network with attention mechanism.

[0045] The invention will appear more clearly on reading the description which follows, given solely by way of non-limiting example, and made with reference to the drawings in which:

[0046] - [Fig.l] [Fig.l] is a schematic view of an electronic estimation device of the pitch of the human voice according to the present invention;

[0047] - [Fig.2] [Fig.2] is a schematic view of an example of the architecture of the device for estimating the pitch of the human voice according to the present invention;

[0048] - [Fig.3] [Fig.3] is a flowchart of a pitch estimation method of the human voice according to one embodiment of the present invention.

[0049] The electronic device for estimating the pitch of the human voice of an audio signal received at input according to the present invention is illustrated in [Fig.l].

[0050] As can be seen in the embodiment of [Fig.l], the electronic device for estimating the pitch of the human voice of an audio signal received at the input comprises a decomposition module 12 configured to decompose said signal into frames of predetermined finite size connected at the output to the input a calculation module 14 configured to calculate, for each frame, an autocorrelation of said frame and a spectral transformation of said frame.

[0051] Said calculation module 14 is connected at output to the input of a module 16 for supplying, the pair of results of said calculation, at input to a neural network 18 with attention mechanism, previously trained, and configured to supply at output the probability that the shift of the autocorrelation corresponds to the period of the pitch of said voice of the audio signal received at input.

[0052] Said neural network with attention mechanism 18 is connected at output to the input of an estimation module 20 configured to estimate said pitch from said probability provided at output by said neural network with attention mechanism 18 as represented in [Fig.l].

[0053] As an optional addition, the electronic device for estimating the pitch of the human voice of an audio signal received at the input comprises a correction module 22 whose input is connected to the output of said neural network with 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 capable of taking into account the evolution of said pitch over several successive frames.

[0054] In the example of [Fig.l], the electronic device for estimating the pitch of the human voice of an audio signal received at the input comprises an information processing unit 24 formed for example of a memory 26 and a processor 28 associated with the memory 26.

[0055] In the example of [Fig.l], the decomposition module, the calculation module and the supply module, the neural network and the estimation module, as well as the optional addition of the correction module, are each implemented in the form of software, or a software brick, executable by the processor. The memory of the electronic device for estimating the pitch of the human voice of an audio signal received at the input is then capable of storing decomposition software, calculation software, supply software, software for implementing a neural network and estimation software, as well as the optional addition of correction software. The processor is then capable of executing each of the software among the decomposition software, the calculation software, the supply software, the software for implementing a neural network and the estimation software as well as the optional addition of the correction software.

[0056] In a variant not shown, the decomposition module, the calculation module, the supply 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 (Field Programmable Gate Array), an integrated circuit, such as an ASIC (Application Specific Interconnection Module). Integrated Circuit), one or more GPU(s) (Graphics Processing Unit), or one or more accelerators dedicated to artificial intelligence (AI).

[0057] When the electronic device for estimating the pitch of the human voice of an audio signal received at the input is produced in the form of one or more software programs, that is to say in the form of a computer program, also called a computer program product, it is also capable of being recorded on a medium, not shown, that is readable by a computer. The computer-readable medium is, for example, a medium capable of storing electronic instructions and of being coupled to a bus of a computer system. For 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.

[0058] Some of these modules and, where applicable, their implementation options are detailed in more detail below.

[0059] More precisely, the decomposition module 12 is capable of decomposing the input audio signal into frames of predetermined finite size, for example 22.5ms, 67.5ms (i.e. 3x22.5ms), 90ms, 20ms, 30ms, etc.

[0060] The calculation module 14 subjects each frame of finite size to two independent transformations.

[0061] Indeed, the calculation module 14 is, as indicated previously, capable of 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.

[0062] As an optional addition, said spectral transformation is a spectral transformation belonging to the group comprising at least:

[0063] - a constant-Q transform CQT (from the English Constant-Q transform);

[0064] - a discrete cosine transform suitable for providing cepstral coefficients at Mel frequency MFCC (from the English Mel-frequency cepstral coefficients).

[0065] - a transform based on the amplitude of a fast Fourier transform FFT (this is the English Fast Fourier Transform);

[0066] - etc.

[0067] Preferably, the constant Q transform is used because it has been experimentally proven to be more efficient than the transform based on the amplitude of a fast Fourier transform FFT.

[0068] It should be noted that the spectral transformation of the frame is defined before the training phase of the neural network 18.

[0069] The supply module 16 is configured to simultaneously supply the pair of results delivered by the calculation module 14 as input to the neural network 18, said pair comprising the result A of autocorrelation of said frame and the result S of said spectral transformation.

[0070] In particular, to form said pair, the supply module 16 is capable of concatenating within 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 learning is arbitrary, the autocorrelation result of said frame being able to be before or after the result of said spectral transformation, but the order of concatenation implemented during learning fixes that implemented later during inference.

[0071] Via said input pair, the results A and S of the two transformations are then processed simultaneously by the neural network 18 which, from a self-attention mechanism, will provide the probability p ( T | £ A) that the lag T of the autocorrelation A corresponds to the period of the voice pitch.

[0072] The probability P ( r | £ A) is a conditional probability knowing both the results S and A of the two transformations, respectively of the spectral transformation and of the autocorrelation of the frame.

[0073] Subsequently, it is considered that the neural network 18 comprises an ordered succession of layers of neurons, each of which takes its inputs from the outputs of the previous layer.

[0074] More precisely, each layer comprises neurons taking their inputs from the outputs of the neurons of the previous layer, or from the input variables for the first layer.

[0075] Alternatively, more complex neural network structures can be envisaged with a layer that can be connected to a layer further away than the immediately preceding layer.

[0076] Each neuron is also associated with an operation, i.e. a type of processing, to be carried out by said neuron within the corresponding processing layer.

[0077] 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 connection between two neurons. It is often a real number, which takes both positive and negative values. In some cases, the synaptic weight is a complex number.

[0078] Each neuron is capable of performing a weighted sum of the value(s) received from the neurons of the previous layer, each value then being multiplied by the respective synaptic weight of each synapse, or connection, between said neuron and the neurons of the previous layer, then applying an activation function, typically a non-linear function, to said weighted sum, and delivering as output to said neuron, in particular to the neurons of the following layer connected to it, the value resulting from the application of the activation function. The activation function makes it possible to introduce non-linearity into the processing carried out by each neuron. The sigmoid function, the hyperbolic tangent function, the Heaviside function are examples of activation functions.

[0079] As an optional addition, each neuron is also capable of applying, in addition, a multiplicative factor, also called 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 from the activation function.

[0080] 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 previous layer.

[0081] Such a type of layer is more often referred to as “fully connected” and sometimes referred to as “dense layer”.

[0082] As an optional addition, the neural network 18 is based on a neural network architecture used in the field of automatic language processing, “pointer networks” such as those introduced in particular by O. Vinyals et al. in the article “Pointer Networks” of June 9, 2015.

[0083] The use of a “Pointer Networks” type architecture in the present field of pitch detection is advantageous because it allows saving memory thanks to a limited number of internal parameters, in numerical complexity, with a limited number of mathematical operations.

[0084] Experimentally, such gains have made it possible to implement the solution according to the present invention in an extremely constrained hardware environment with a power consumption of less than 0.05W.

[0085] A training phase of the neural network 18, before implementing the solution, makes it possible to make the solution robust to acoustic noise. More precisely, said neural network, preferably of the “pointer network” 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.

[0086] As an optional addition, the latent space of said neural network is capable of being, beforehand, partitioned according to a partitioning (also called clustering from the English data clustering).

[0087] According to a first variant of this optional complement, said partitioning is a partitioning via an unsupervised method such as a partitioning in k- averages (or k-means in English) or partitioning via Gaussian mixture models (GMM).

[0088] Such prior partitioning of the latent space of the neural network 18 makes it possible to detect transitions from voiced frames (i.e. including human voice) to unvoiced frames (i.e. not including any human voice).

[0089] More precisely, the principle is to force a partition into two (after having verified it empirically on the data) and each partition corresponds to voiced or unvoiced.

[0090] More precisely, the determination of the unvoiced / voiced transition is said to be semi-supervised because it is obtained both using partitioning (clustering) according to an unsupervised method and using a supervised training phase of said neural network 18.

[0091] 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.

[0092] According to a first embodiment, 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 provided as output by said neural network with attention mechanism 18.

[0093] In the context of this first embodiment (i.e. without correction processing / processing before estimation), the estimation module 20 is capable of selecting the autocorrelation shift 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 shift.

[0094] 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 capable of taking into account the evolution of said pitch over several successive frames.

[0095] 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.

[0096] [Fig.2] is a schematic view of a detailed example of the architecture of the device 10 for estimating the pitch of the human voice according to the present invention.

[0097] The modules identical to [Fig.l] are reported in [Fig.2] with identical numbering.

[0098] 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.

[0099] The calculation module 14 comprises, as illustrated by [Fig.2] two calculation tools 30 and 31 implemented in parallel according to the example of [Fig.2].

[0100] The first calculation tool 30 is configured to calculate the autocorrelation A of each frame provided by the decomposition module 12.

[0101] The second calculation tool 31 is configured to apply a spectral transformation to each frame provided by the decomposition module 12. More precisely, according to the example of [Fig.2], the spectral transformation 32 implemented by the calculation tool 31 and defined before training the neural network 18 corresponds to a constant Q transformation CQT.

[0102] As an optional addition, as illustrated by the example of [Fig.2], the tool 31 for calculating the spectral transformation further comprises a convolutional neural network 34, previously trained, and capable of advantageously implementing a formatting F, capable of making the result S provided by said tool 31 for calculating said spectral transformation concatenable with the result A provided by said tool 30 for calculating the autocorrelation of said frame.

[0103] A convolutional neural network is also sometimes called a convolutional neural network or by the acronym CNN which refers to the English term “Convolutional Neural Networks”.

[0104] In a convolutional neural network, each neuron in the same layer has 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, a "kernel" in reference to the corresponding English term.

[0105] More precisely, said convolutional neural network 34 comprises, according to the architectural example of [Fig.2], a dense layer, Nc convolutional layers called ConvB, with Ncun integer, a formatting layer F (flatten in English) and a dense layer.

[0106] 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. Note that once this order is fixed during learning, the same order must be used during inference.

[0107] The results A and S of the calculation module 14 are provided respectively, for the same frame, by the two independent calculation tools 30 and 31 at the input of the module. 16 of supply configured to simultaneously supply, via concatenation, during said supply the pair of results delivered by the calculation module 14 as input to the neural network 18 with attention mechanism.

[0108] More precisely, according to the example of [Fig.2], the neural network 18 is preferably 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 closed recurrent unit GRU (from the English Gate Recurrent Unit).

[0109] 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 shift of the autocorrelation A and the period of the pitch, and this knowing the frequency representation S of the frame.

[0110] At the output of this stage 38 of self attention the similarity is transformed via a probabilistic model, in particular by means of the mathematical function softmax 40 into the probability 42 (r|S. A) 9ue 'c lag r (in English lag) of the autocorrelation A corresponds to the period of the Pitch of the voice.

[0111] This probability 42 is then used, according to the example of [Fig.2], directly by the pitch estimation module 20.

[0112] In other words, the example of [Fig.2] illustrates the first embodiment mentioned above, 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 provided as output by said neural network with attention mechanism 18.

[0113] In the context of this first embodiment (i.e. without correction processing / processing before estimation), the estimation module 20 is capable of selecting the autocorrelation shift 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 shift.

[0114] An example embodiment of the operation of such a device 10 according to the present invention is described below in relation to [Fig. 3].

[0115] More precisely, the method 50 for estimating the pitch of a human voice from an audio signal received as input firstly comprises a first step 52 of decomposing said signal into TR frames of predetermined finite size.

[0116] Then, the method 50 comprises a step 54 of calculating two independent transformations of each frame resulting from the decomposition step 52.

[0117] More precisely, said calculation step 54 comprises two independent calculation sub-steps 56 and 58, namely a sub-step 56 of calculation C_A of an autocorrelation of said frame, and a sub-step 58 of calculation C_T_S of a spectral transformation of said frame.

[0118] As an optional addition, said spectral transformation 68 is a spectral transformation belonging to the group comprising at least:

[0119] - a constant Q transform;

[0120] - a discrete cosine transform suitable for providing cepstral coefficients at Mel frequency;

[0121] - a transform based on the amplitude of a fast Fourier transform.

[0122] These two sub-steps 56 and 58 are suitable for being implemented in parallel, as illustrated by [Fig.3], or successively in any order.

[0123] According to an optional aspect, as indicated previously, 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 to form, via concatenation during said provision, said pair of results provided as input to said neural network with attention mechanism.

[0124] Then the method 50 comprises a step 62 of providing the pair of results of said calculation as input to a step 64 of inference phase P_I of the neural network with attention mechanism 18, said pair being formed in particular via concatenation.

[0125] As illustrated by [Fig.3], the neural network with attention mechanism 18 used during the inference phase 64 is configured to provide as output the probability 42 that the shift of the autocorrelation corresponds to the period of the pitch of said voice of the audio signal received as input, and was obtained at the end of a prior step 66 corresponding to a training phase P_E of said neural network with attention mechanism.

[0126] 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 partitioning in k-means or a partitioning via Gaussian mixture models GMM (from the English Gaussian Mixture Models).

[0127] Preferably, the second aforementioned 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 correction of 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.

[0128] The corrected probability resulting from this optional correction step 72 is then used during a step 76 of estimating said pitch.

[0129] As an alternative according to arrow 78, according to the first embodiment mentioned above, the probability 42 resulting from the inference phase 64 is directly used (i.e. without subsequent correction processing of the Viterbi algorithm type) during the step 76 of estimating said pitch.

[0130] Those 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 each other to generate new embodiments of the invention.

[0131] The present invention thus allows use in low-bitrate vocoders, for example, due to the cross-use of autocorrelation information and spectral information (any, but defined during the definition of the neural network), in other words due to the use of two types of information on the frame.

[0132] Furthermore, the use of a specific neural network architecture of the “Pointer Networks” type makes it possible to have very small neural networks compared to the state of the art, which allows for low complexity and a rapid inference phase that requires little resource.

[0133] Furthermore, as indicated, the learning phase (i.e. training) of the neural network is capable of being adapted, via semi-supervised training, to operate in a noisy environment, which makes it possible to go beyond the mathematical definition of Pitch reduced to the fundamental frequency of the voice.

[0134] The solution proposed according to the present invention is thus robust to noise thanks to its training phase because the implementation of strong hypotheses on the properties of the noise is avoided.

Claims

Claims

1. 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: - decomposition (52) of said signal into frames of predetermined finite size; - for each frame, - calculation (54): - of an autocorrelation of said frame; and - of a spectral transformation of said frame; - supplying (62) the pair of results of said calculation as input to a neural network with 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, estimation (76) of said pitch.

2. Estimation method (50) according to claim 1, wherein said spectral transformation is a spectral transformation belonging to the group comprising at least: - a constant Q transform; - a discrete cosine transform capable of providing cepstral coefficients at Mel frequency; - 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 partitioning (68) the latent space of said neural network.

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

5. An estimation method (50) according to any one of the preceding claims wherein the autocorrelation offset corresponding to the maximum of said probability is 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 offset.

6. 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 capable of taking into account the evolution of said pitch over several successive frames.

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

8. Estimation method (50) according to any one of the preceding claims, wherein said calculation of said spectral transformation 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 to form, via concatenation during said provision, said pair of results provided as input to said neural network with attention mechanism.

9. A computer program comprising software instructions which, when executed by a computer, implement a method of estimating the pitch of a voice of an audio signal received as input according to any one of the preceding claims.

10. Electronic device (10) for estimating the pitch of a human voice of an audio signal received as 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 transformation of said frame; - a module (16) for providing the pair of results of said calculation as input to a neural network with attention mechanism, previously trained, and configured to provide 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; - said neural network with attention mechanism; - an estimation module (20) configured to estimate said pitch from said probability provided as output by said neural network with attention mechanism.

Images