Method for verifying a speaker by a neural network, associated device and motor vehicle.
The method enhances speaker verification in motor vehicles by using a convolutional neural network to process voice recordings, emphasizing relevant frequencies, thereby improving reliability and accuracy in identifying authorized speakers.
Patent Information
- Application Number
- FR2024000209
- Authority / Receiving Office
- FR · FR
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-01-10
- Publication Date
- 2025-07-11
- Estimated Expiration
- 2044-01-10
AI Technical Summary
Existing speaker verification systems in motor vehicles lack reliability, particularly in distinguishing between different speakers based on their voices.
A method using a convolutional neural network to process voice recordings, employing attention operations and matrix transformations to enhance the importance of relevant sound frequencies, followed by comparison with a reference output to verify speaker identity.
Improves the reliability of speaker verification by focusing on critical sound frequencies, ensuring accurate identification of authorized speakers.
Smart Images
Figure 00000000_0000_ABST
Abstract
Description
Title of the invention: Method for verifying a speaker by a neural network, associated device and motor vehicle.
[0001] The invention relates to motor vehicles and speaker verification by a convolutional neural network.
[0002] In particular, there is a need to improve the reliability of such verification.
[0003] For this purpose, the invention relates to a method for verifying (an identity) of a speaker, implemented in a motor vehicle, comprising the following steps: • Obtaining (in other words: receiving or determining) a verification input representative of (in other words: obtained from) a digitized sound recording (obtained from) the speaker's voice, in the form of a matrix of H rows and W columns, H (for example equal to 80) and W (for example equal to 300) being even integers greater than 8 (and for example less than 500), the H rows being representative of different (and / or disjoint) sound frequencies (in other words: sound frequency ranges) (in other words: each of the H rows being representative of the sound recording at a different (or disjoint) sound frequency (in other words: in a sound frequency range) from the sound frequencies (in other words: sound frequency ranges) of which the other H1 rows are representative), the W columns being representative of instants (in other words:of different (and / or disjoint) time ranges (in other words: each of the W columns being representative of the sound recording at a time (in other words: in a time range) different (or disjoint) from the times (in other words: time ranges) of which the other Wl columns are representative),; • Producing a verification output from the verification input by a first electronic convolutional neural network comprising an upstream part and a downstream part, the production of the verification output comprising the following steps: • Determination of a working matrix comprising H rows and W columns, and C channels, C being an integer strictly greater than 1, including an implementation, by the upstream party: • From an initial convolution operation Conv0, by an initial convolution layer of the upstream part, taking as input, the verification input, an output of the operation of initial convolution Conv0 (in other words; of the initial convolution layer) having C channels, H rows and W columns (the person skilled in the art knows and knows how to implement such an operation and such a layer), then of an initial attention operation ATT0, by an initial attention module of the upstream part, the initial attention operation ATT0 consisting of: • A step of obtaining an initial attention line vector (in other words: an attention vector) by initial neural layers of the initial attention module from initial data obtained from an output of the initial convolution operation Conv0 (in other words; of the initial convolution layer) (for example, an initial input vector of H values, each vth value of the initial input vector being equal to the average of the values of the vth lines of the C channels of the output of the operation (in other words: of the layer) of convolution (v between H and 1)), then • A multiplication of the initial attention row vector of H values, by each of the C channels of the output of the initial convolution operation Conv0 (in other words; of the initial convolution layer), (ie: a first result of the initial attention operation ATT0 being a matrix of H rows, W columns, and C channels, each channel of the C channels being a result of the multiplication), Where: • Th,w,c — ATTo (Conv0 (EVHjW)), • THjWjC is the working matrix, • evh.w is the verification input, • Then, production of an intermediate output, from the working matrix, by a first part of the downstream part, the intermediate output comprising H rows and W columns, and C channels, then • Production of the verification output from the intermediate output by a second part of the downstream part, Then, comparing a distance between the verification output and an output reference with a threshold, then, • Control of a component (for example: a steering system, a display screen, a propulsion motor, brakes, lights, a horn, a microprocessor, opening locks, a motor vehicle ignition switch) of the motor vehicle, provided that the distance between the verification output and the reference output is less than the threshold, i.e. that the identity of the speaker is verified.
[0004] According to the invention, the attention operation thus makes it possible to improve the reliability of the verification by giving more importance to the frequencies which are relevant for the verification of the speaker.
[0005] The distance is for example a Euclidean distance, or alternatively, another type of distance type.
[0006] The method may comprise, for example, receiving a sound recording from a microphone and obtaining the verification input by digitizing the sound recording (for example by a signal processing module of the microprocessor), in a known manner.
[0007] The values of the verification input EVH,w are for example on the mel scale (alternatively, in decibels).
[0008] The reference output can be obtained at the output of the first electronic convolutional neural network, following a step of implementing the first electronic convolutional neural network (for example, during a speaker enrollment step), with as input, a verification input representative of a digitalized sound recording of the voice of a speaker, for example the owner or usual driver of the motor vehicle.
[0009] According to one embodiment, the production of the intermediate output, from the working matrix, comprises the following steps, implemented by the first part: • Expansion steps in which the first part increases a number of channels of expanded matrices, from the working matrix, by convolution operations, while decreasing the number of rows and columns of the expanded matrices, by subsampling operations, then • Contraction steps in which the first part decreases a number of channels of contracted matrices, by deconvolution operations, from the expanded matrices, while increasing the number of rows and columns of the contracted matrices, by oversampling operations, the intermediate output being obtained from the contracted matrices.
[0010] An example of such an architecture called “U-net” is for example described in the following article: • Olaf Ronneberger, Philipp Fischer, and Thomas Brox. “U-net: Convolutional networks for biomedical image segmentation”, 2015.
[0011] Alternatively, other architectures may be used, with or without expansion or contraction steps, such as, for example, the architecture described in the article below: • Yafeng Chen, Siqi Zheng, Hui Wang, Luyao Cheng, Qian Chen, Jiajun Qi, “An Enhanced Res2Net with Local and Global Feature Fusion for Speaker Verification,” 2023.
[0012] According to one embodiment, the expansion steps comprise (in other words: the method comprises), for any integer n such that 0 <n<N, N étant un entier (prédéterminé) tel que 2N<H et 2N<W, la génération d’une matrice EH / 2n W / 2n2nc, par une mise en œuvre, par la première partie :
[0013] From an nth convolution operation Convn, implemented by an nth convolution layer of the first part, an output of the nth convolution operation Convn (or of the nth convolution layer) comprising 2nC (ie: 2n*C) channels, H / 2n rows, W / 2n columns (the person skilled in the art knows and knows how to implement such an operation and such a layer), then
[0014] From an nth attention operation ATTn, by an nth attention module of the first part, the nth attention operation ATTn consisting of: • A step of obtaining an nth attention line vector (in other words; an nth attention vector), of H / 2"1 values, by nth neural layers of the nth attention module from data obtained from the output of the nth initial convolution operation Conv0 (in other words: of the nth convolution layer) (for example, an input vector of H / 2"1 values, each vth value of the input vector being equal to the average of the values of the vth lines of the 2nC channels of the output of the nth initial convolution operation Convn (in other words: of the nth convolution layer (v between H and 1)), then • A multiplication of the nth attention row vector, by each of the 2nC channels of the output of the nth convolution operation Convn (in other words; of the nth convolution layer), (ie: a first result of the nth attention operation ATTn being a matrix of H / 2"1 rows, W / 2"1 columns, and 2nC channels, each channel of the C channels being a result of the multiplication), • A MAXPOOL sub-sampling operation, implemented by a sub-sampling layer (in English “a pooling layer”) of the first part, an output of the MAXPOOL subsampling operation being a matrix having H / 2n rows, W / 2" columns and 2nC channels (those skilled in the art know several MAXPOOL subsampling operations, for example obtaining, at the output of MAXPOOL, the maximum or an average of (neighboring) values in an input of MAXPOOL), and • Or: • E^,w / 2yc = MAXPOOL(ATTn (Convn (Eh / ^w^^'c)))^ * EHjWjC = THjWjC, • We denote by M1JjK a matrix M of I rows, J columns and K channels, whatever let the integers I, J and K and the matrix M (for example M=E, I=H / 2N, J=W / 2N, K=C*2n).
[0015] Of course, as a variant, other different expansion steps are conceivable.
[0016] According to one embodiment, the contraction steps comprise (in other words: the method comprises), for any integer m such that 0 <m<N, la génération d’une matrice SIh / 2™ w / 2m 2mc par une mise en œuvre, par la première partie: • An EXP upsampling operation implemented by an upsampling layer of the first part, an output of the EXP upsampling operation being a matrix having H / 2m rows, W / 2m columns and 2m+1C channels (those skilled in the art know several EXP downsampling operations, for example adding, at the output of EXP, zeros or duplicating the values at the input of EXP), then • Of a mth deconvolution operation DConvm implemented by a mth deconvolution layer of the first part, an output of the mth deconvolution operation DConvm (or of the mth deconvolution layer) comprising 2mC (ie: 2m*C) channels (the person skilled in the art knows and knows how to implement such an operation and such a layer), the mth deconvolution operation DConvm (or of the mth deconvolution layer) taking as input two matrices (comprising a matrix of H / 2m rows, W / 2m columns, and 2m+1C channels and a matrix of H / 2m rows, W / 2m columns, and 2mC channels), • Or : • SIH / 2mw / 2m2mc = DConvm (EH / 2m w / 2m2mc, EXP(SIH / 2mn W / 2m+1.2(m+1)c)), and * SIh / 2N W / 2N2NC = EH / 2N W / 2N 2Nc, • We denote by M1jJjK a matrix M of I rows, J columns and K channels, whatever the integers I, J and K and the matrix M (for example M=E, I=H / 2N, J=W / 2N, K=C*2N).
[0017] Of course, as a variant, the contraction steps may be different, for example example to combine the matrices EH / 2n W / 2n 2nc in a single operation, in a different order or in a different way. Also, attention steps (e.g., on frequencies) could be implemented during the contraction steps.
[0018] According to a first embodiment, the verification output is a matrix consisting of a column of means, of H values, and a column of variances, of H values, each pth value, p being an integer such that l <p<H, parmi les H valeurs de la colonne de moyennes, étant une moyenne (mathématique) des valeurs des pième lignes des C canaux de la sortie intermédiaire, chaque qième valeur, q étant un entier tel que l<q<H , pami les H valeurs de la colonne de variances étant une variance (mathématique) des valeurs des qième lignes des C canaux de la sortie intermédiaire.
[0019] Of course, as a variant, the verification output may have a different form.
[0020] According to one embodiment, the method comprises, prior to receiving the verification input, a step of training the first electronic convolutional neural network including (in other words: comprising) a step of training a second electronic convolutional neural network, the second electronic convolutional neural network consisting of the first electronic convolutional neural network, a training output layer, at the output of the second electronic convolutional neural network, and a fully (in other words: complement) connected layer connecting the verification output to the training output, the step of training the second electronic convolutional neural network (and therefore, the first electronic convolutional neural network) being implemented from a plurality of training matrices,as input to the second electronic convolutional neural network (i.e. or therefore, as input to the first electronic convolutional neural network) comprising H rows and W columns, representative of digitized sound recordings of voices of a plurality of u speakers (i.e.: each training matrix of the plurality of training matrices being representative of a digitized sound recording of the voice of a second speaker among the plurality of u speakers), labeled by a plurality of speaker identification matrices (in other words: each training matrix of the plurality of training matrices is labeled by a speaker identification matrix among a plurality of u speaker identification matrices) in one-to-one correspondence with the plurality of u speakers, u being an integer.
[0021] In a known and conventional manner, during the training of the second electronic convolutional neural network, the weights of the convolution and deconvolution layers of the neural network, and of the fully connected layer are modified, for example by gradient backpropagation.
[0022] For example, according to the first embodiment, each matrix of the plurality of speaker identification matrices comprises u rows and 2 columns, and comprises a predetermined value (for example equal to 1) in its 2 columns at a different tth row for each speaker of the plurality of u speakers.
[0023] For example, the second neural network can be trained from a dataset called “VoxCeleb 1” comprising 153516 sound recordings from 1251 speakers.
[0024] The invention also relates to a computer program comprising instructions executable by a microprocessor or a microcontroller or a computer, to implement the steps of the method according to the invention, when it is executed by the microprocessor or the microcontroller or the computer.
[0025] The method according to the invention can be implemented by an electronic device (or a motor vehicle). The invention therefore also relates to an electronic device (or a motor vehicle) configured to implement the steps of the method according to the invention, as well as a motor vehicle comprising the electronic device, the electronic device comprising the first electronic convolutional neural network.
[0026] When the method is implemented by an electronic device, it is understood, for example, that an element of a neural network, for example a convolution layer or a deconvolution layer implements an operation, by the fact that the element is part of the electronic device and performs the operation.
[0027] The characteristics and advantages of the computer program, the electronic device, and the vehicle are identical to those of the method according to the invention (without it being necessary to repeat them here).
[0028] When the electronic device, the motor vehicle, (or another element) is "configured to" (or "capable of") performing or implementing a step or operation, this implies, for example, that the element comprises means for performing the step or operation. The means preferably comprise electronic means, for example a computer program, data in memory, specialized electronic circuits, wired or wireless connections, a microprocessor and / or a microcontroller.
[0029] Other characteristics and advantages of the present invention will appear more clearly on reading the detailed description which follows, comprising embodiments of the invention given as non-limiting examples and illustrated by the appended drawings, in which: • [Fig.l] represents an electronic device and a motor vehicle, according to an embodiment of the invention, in top view, • [Fig.2] to [Fig.4] represent a neural network of the electronic device of [Fig.l] and elements of this neural network. • [Fig.5] represents an implementation of the method according to the invention, according to a exemplary embodiment, by the electronic device and the motor vehicle of [Fig.l].
[0030] In [Fig.l], certain elements are, of course, seen through transparency.
[0031] Detailed description of an exemplary embodiment of the invention, with reference to Figures 1 to 5.
[0032] In the following, we note: MAp a matrix MA of I rows and J columns, whatever the integers I, J and the matrix MA, MB! j K an MB matrix of I rows, J columns and K channels, whatever the integers I, J and K and the MB matrix.
[0033] [Fig.l] represents a vehicle 100 which is a motor vehicle. The vehicle 100 comprises a microprocessor 110 to which the following components of the motor vehicle 100 are connected: A 120 propulsion engine, A 130 direction, 140 lights, A 150 display screen, A 160 microphone, 170 brakes, 180, 190 opening locks.
[0034] The microprocessor 110 also comprises the neural network 111_100.
[0035] [Fig.5], in step S00, the neural network 200 [Fig.2] is trained on a computer (not shown) outside the vehicle 100.
[0036] The neural network 200 comprises: A neural network 111_100, comprising, as output, a verification output 122_SV8o,2, A training output layer 200_OUT, at the output of the neural network 200, which is a matrix comprising 1251 rows and 2 columns. A fully connected layer 200_FULL connecting the verification output 122_SV8o,2 to the training output 200_OUT.
[0037] The neural network 200 and the neural network 111_100 are convolutional neural networks.
[0038] Step S00 is a supervised training implemented from training matrices, as input to the neural network 200 (and therefore to the neural network . 111_100) obtained from the data set called “VoxCeleb 1” comprising 153516 sound recordings of 1251 speakers. The training matrices are labeled by matrices representative of the speaker whose voice is present on each recording. In a known and conventional manner, during training, the weights of the convolution and deconvolution layers of the neural network .111_100 and the fully connected layer 200_FULL are modified.
[0039] The neural network . 111_100 can then be installed, in step S10, in the microprocessor 110. During step S10, a reference matrix 122_SR80,2 can also be recorded in the memory of the microcontroller 110. The reference matrix 122_SR8o,2 can be obtained at the output of the neural network 111_100 following a step of implementing the neural network 111_100, with as input, a verification input representative of a digitalized sound recording of the voice of a speaker, for example the owner or usual driver of the motor vehicle 100.
[0040] In step S20, before authorizing the sending of a command, by the microcontroller 110, to one of the components of the motor vehicle 100 above, the microcontroller 110 receives a sound recording, coming from the microphone 160, and obtains the verification input HO_EV8O,i6o by digitizing the sound recording (for example by a signal processing module of the microprocessor 110).
[0041] In step S30, the convolution layer 110_Conv of the initialization module 110_Init, represented [Fig.3], of the part 111_110 of the neural network 111_100 implements a convolution operation Conv0, taking as input, the verification input HO_EV8o,i6o, an output 110_Out80,i6o,i6 of the initial convolution operation Convo having 16 channels, 80 rows and 160 columns.
[0042] In step S40, the attention module 110_Att of the initialization module 110_Init implements an attention operation ATT0.
[0043] During this attention operation ATT0, neural layers 110_Att_couch obtain an attention row vector 110_vectAtt of 80 values from a vector 110_vectInit of 80 values. Each vth value of the vector 110_vectInit is equal to the average of the values of the vth rows of the 16 channels of the output HO_Out80>i6o,i6- Then, the initial attention row vector 110_vectAtt is multiplied by each of the 16 channels of the output 110_Out80ji6o,i6- The result of this multiplication is the matrix 110_T 8o,i6o,i6- Thus, 110_T80ji60,i6 = ATT0 (Conv0 (HO_EV8o,i6o))-
[0044] In step S50, the module 121augi of the part 111_121, included in the part 111_120 of the neural network 111_100, produces the matrix 121_E4Oj8o,32 from the matrix 1 10_T80j160,16-
[0045] For this: • The convolution layer 121_Convi of the 121augi module, shown [Fig.4], implements a Convi convolution operation, taking as input the matrix 110_T80,i60,i6, and producing an output 121_Out80,i6o,32- • The 121_Atti attention module of the 121augi module then implements an ATTi attention operation.
[0046] During this attention operation ATTb of the neuronal layers 121_Att_couchi ob hold an attention row vector 121_vectAtti of 80 values from a 121_vectlniti of 80 values. Each vth value of the vector 121_vectlniti is equal to the average of the values of the vth rows of the 32 channels of the output 121_Out80ji6o,32- Then, the attention row vector 121_vectAtti is multiplied by each of the 32 channels of the output 121_Out80ji6o,32- A subsampling layer 121_POOL of the 121augi module then implements a MAXPOOL subsampling operation with the result of this multiplication as input and the matrix 121_E40,80,32 as output. Thus, 121_E4O,8o,32 = MAXPOOLCATT! (ConV1 (HO_T8O,16o,i6))).
[0047] The values of the matrix at the output of the MAXPOOL operation are for example the maximum of 4 neighboring values in the input of this operation, but other variants of sub-sampling operation, known to those skilled in the art, are of course conceivable.
[0048] In step S60, the module 121aug2 of the part 111_121, included in the part 111_120 of the neural network of the neural network 111_100, produces the matrix 121_E2o,4o,64 from the matrix 121_E40,80,32 - the module 121aug2 is similar to the module 121aug2. Therefore, it will not be detailed here.
[0049] In step S70, an oversampling layer 121_expi of the part 111_121 implements an oversampling operation EXP, taking, as input, the matrix 121_E20,40,64. The oversampling operation 121_expi produces as output a matrix which, together with the matrix 121_E40,80,32 is, as input to a deconvolution operation DConvi, implemented by the deconvolution layer 121_DConvi, in step S80, and producing the matrix 121_Sl4o,8o,32- Thus I2IJSI40,80,32 = DConvi (121_E 40,80,32, EXP(121_E20j40,64)-
[0050] In the same way, in steps S70 and S90, the matrix 121_SI8O,i6o,i6 is obtained from the matrix 121_SI40,80,32 by the oversampling layer 121_exp0, then by the deconvolution layer 121_DConv0.
[0051] In step S100, the part 111_122 of the neural network 111_100 produces, from the matrix 121_SI8O,i6o,i6, a verification matrix 122_SV80,2 consisting of a column of means, of 80 values, and a column of variances, of 80 values, each pth value, p being an integer such that l <p<80, parmi les 80 valeurs de la colonne moyennes, étant une moyenne des pième lignes 16 canaux matrice 121_si8o,i6o,i6, chaque qième valeur q un entier tel que l<q<80 , pami variances variance (mathématique) 121_si8o,i6o,i6-
[0052] In step S1 10, the microprocessor 110 determines whether a distance between the verification matrix 122_SV80,2 and the reference matrix 122_SR80,2 is less than a threshold.
[0053] In step S120, if this is the case, the microprocessor 110 sends said command to said one of the components of the motor vehicle 100.
Claims
Claims
1. A method of verifying a speaker, implemented in a motor vehicle (100), comprising the following steps: • Reception (S20) of a verification input (1 lO_EV8O,i6o) representative of a digitalized sound recording of a speaker's voice, in the form of a matrix of H rows and W columns, H and W being even integers greater than 8, the H rows being representative of different sound frequencies, the W columns being representative of different instants, • Production of a verification output (122_SV80>2) from the verification input (1 1O_EV8Oji6o) by a first electronic convolutional neural network (lll_100) comprising an upstream part (111_110) and a downstream part (111_120), the production of the verification output (122_SV80>2) comprising the following steps: • Determination of a working matrix (110_T80>i60,i6) comprising H rows and W columns, and C channels, C being an integer strictly greater than 1, comprising an implementation, by the upstream part (111_110): • Of an initial convolution operation ConvO (S30), by an initial convolution layer (110_Conv) of the upstream part (111_110), taking as input, the verification input (110_EV80,160), an output (121_0ut80,160,16) of the initial convolution operation ConvO having C channels, H rows, and W columns, then • From an initial attention operation ATTo (S40), by an initial attention module (110_Att) of the upstream part (111_110), the initial attention operation ATT0 consisting of: • A step of obtaining an initial attention line vector (110_vectAtt), of H values, by initial neural layers (110_Att_couch) of the module initial attention (110_Att) from an initial data (110_vectInit) obtained from the output (110_Out8o,i6o,i6) of the initial convolution operation Conv0, then • A multiplication of the initial attention line vector (110_vectAtt), by each of the C channels of the output (110_Out80,160,1e) of the initial convolution operation Conv0, where: • Th,w,c= ATTo (Conv0 (EV h,w)X • THjWjC is the matrix of work (11O_T8Oji6o,i6), • EVh,w is the verification input (llO_EV8O,i6o), • Production of an intermediate output (121_SI8O,i6o,i6) from the working matrix (110_T80ji60,i6), by a first part (111_121) of the downstream part (lll_120), the intermediate output (121_SI8O,i6o,i6) comprising H rows and W columns, and C channels, • Production (S 100) of the verification output (122_SV80,2) from the intermediate output (121_SI80,160,1e) by a second part (111_122) of the downstream part (111_120), • Comparison (SI 10) of a distance between the verification output (122_SV80,2) and a reference output (122_SR80,2) with a threshold, • Control (S 120) of a component of the motor vehicle (100) conditioned on the distance between the verification output (122_SV80,2) and the reference output being less than the threshold.
2. Verification method according to the preceding claim in which the production of the intermediate output (121_SI8O,i6o,i6) comprises the following steps, implemented by the first part (111_121): • Expansion stages during which the first part (111_121 ) increases a number of channels, of expanded matrices, by convolution operations, from the working matrix (110_T80,i60,i6), while decreasing a number of rows and columns of the expanded matrices, by downsampling operations, Contraction steps during which the first part (111_121 ) decreases a number of channels, of contracted matrices, by deconvolution operations, from the expanded matrices, while increasing a number of rows and columns of the contracted matrices, by upsampling operations the intermediate output (121_SI8O,i6o,i6) being obtained from the contracted matrices.
3. Verification method according to the preceding claim in which the expansion steps comprise, for any integer n such that 0 <n<N, N étant un entier tel que 2N<H et 2N<W, une génération d’une matrice EH / 2n ,w / 2n2nc ( 121_E40j80,32, 121_E2o,40,64^ par une mise en œuvre, par la première partie (111_121): From an nth convolution operation Convn, implemented by an nth convolution layer (121_Convi) of the first part (111_121), an output (121_Out80>i6o,32) of the nth convolution operation Convn comprising 2nC channels, H / 2n 1 rows, W / 2"1 columns, From an nth attention operation ATTn, by an nth attention module (12l_Atti) of the first part (111_121), the nth attention operation ATTn consisting of: • A step of obtaining an nth attention line vector (121_vectAtti), of H / 2"1 values, by nth neural layers (121_Att_couchi) of the nth attention module (121_Atti) from data obtained from the output (121_Out80>i6o,32) of the nth initial convolution operation Convn • A multiplication of the nth attention line vector (121_vectAtti), by each of the 2nC channels of the output (121_Out80>i6o,32) of the nth convolution operation Convn, From a MAXPOOL subsampling operation, implemented implemented by a subsampling layer (121_POOL) of the first part (111_121 ), an output of the MAXPOOL subsampling operation being a matrix having H / 2n rows, W / 2n columns and 2nC channels, • Or: EH / 2n w / 2n2nc = MAXPOOL(ATTn (Convn (Eh^^1 / 1^))), • EHjWjc = THjWjc, • We denote by M1J>K a matrix M of I rows, J columns and K channels, whatever the integers I, J and K and the matrix M
4. Verification method according to the preceding claim in which the contraction steps comprise, for any integer m such that 0 <m<N, une génération d’une matrice SIH / 2mw / 2m2mc ( 121-8140,80,32) par une mise en œuvre, par la première partie (111_121 ): • An EXP upsampling operation (S70) implemented by an upsampling layer (121_expi, 121_exp0) of the first part (111_121), an output of the EXP upsampling operation being a matrix having H / 2m rows, W / 2m columns and 2m+1C channels, • From a mth DConvm deconvolution operation (S80) implemented by a mth deconvolution layer (121_DConvi, 121_DConv0) of the first part (111_121), an output of the mth DConvm deconvolution operation comprising 2mC channels, the mth DConvm deconvolution operation taking two matrices as input, • Or: • SIH / 2mW / 2m2mC = DConvm (EH / 2mW / 2m2mC, EXP( S IH / 2m+^ / 2^,2^° c)), and * SIh / 2N,W / 2N,2NC = EH / 2N,w / 2N,2NC-
5. A verification method according to any one of the preceding claims wherein the verification output (122_SV8o,2) is a matrix consisting of a column of means, of H values, and a column of variances, of H values, each pth value, p being an integer such that l <p<H, parmi les H valeurs de la colonne de moyennes, étant une moyenne de valeurs de pième lignes des C canaux de la sortie inter- median (121_SI80,100,10), each qth value, q being an integer such that l <q<H, pami les H valeurs de la colonne de variances, étant une variance de valeurs de qième lignes des C canaux de la sortie intermédiaire (121_SI80,160,16).
6. Verification method according to any one of the preceding claims, comprising, prior to receiving (S20) the verification input (1 1O_EV8Oji6o), a step of training the first electronic convolutional neural network (111_100) including a step of training (S00) a second electronic convolutional neural network (200), the second electronic convolutional neural network (200) consisting of the first electronic convolutional neural network (lll_100), a training output layer (200_OUT), at the output of the second electronic convolutional neural network, and a fully connected layer (200_FULL) connecting the verification output (122_SV80>2) to the training output (200_OUT), the step of training (S00) the second electronic convolutional neural network (200) being implemented from a plurality of training matrices,as input to the second electronic convolutional neural network (200) comprising H rows and W columns, representative of digitized sound recordings of voices of a plurality of u speakers, labeled by a plurality of speaker identification matrices in one-to-one correspondence with the plurality of u speakers, u being an integer.,
7. Computer program comprising instructions, executable by a microprocessor or a microcontroller, for implementing the method according to any one of claims 1 to 6, when executed by the microprocessor or the microcontroller.
8. An electronic device (110) configured to implement the steps of the method according to any one of claims 1 to 6, the electronic device comprising a first electronic convolutional neural network.
9. Motor vehicle (100) comprising the electronic device (110) according to the preceding claim.
Citation Information
Patent Citations
Voiceprint identity verification method for automatic sliding door of vehicle
CN116486820A