Speaker recognition method in real sound field environment based on self-supervised learning model

Abstract

Disclosed is a speaker recognition method in a real sound field environment based on a self-supervised learning model. The method includes the steps of: acquiring large-scale unlabelled pure audio dat

Description

TECHNICAL FIELD The present invention relates to the technical field of speaker recognition and deep learning, and particularly, refers to a speaker recognition method in a real sound field environment based on a self-supervised learning model. BACKGROUND Speaker recognition is a task to identify a speaker, comprising speaker identification, speaker confirmation and multi-speaker classification. Everyone's unique sound track structure, pronunciation habit and accent make it possible to identify the speaker through voiceprint features in speech, which is similar to fingerprint, face and iris recognition methods, and the voiceprint recognition method has the advantages of no contact and high recognition precision, thereby being widely used in access control, user authentication, evidence collection, and other aspects. In an ideal sound field environment (the speaker's speech features are not interfered by environmental noise, sensing noise, and other speakers' speeches and echoes), various speaker recognition methods based on a deep neural network have achieved excellent effects in an equal error rate and other key indexes. However, the speaker recognition problem in the case of strong noise and reverberation when a signal-to-noise ratio is close to 0 is rarely studied. In an actual sound field environment, there are often sudden noise signals with high energy and obvious speech reverberation. For example, in a closed conference room, there may be a door opening and closing sound, a coughing sound, a footstep sound, and the like. In addition, in a far-field sound field environment, a lower signal-to-noise ratio between a speech signal and acoustic sensor noise cannot be ignored either, which leads to a sharply reduced speaker recognition effect. Therefore, in a real sound field environment, it is difficult to achieve high-precision speaker identification. SUMMARY Objective of invention: in order to solve the problem that it is difficult to achieve high-precision speaker identification in a strong noise and reverberation environment by a speaker recognition device in the prior art, the present invention provides a 1 speaker recognition method in a real sound field environment based on a self-supervised learning model. Technical solution: a speaker recognition method in a real sound field environment based on a self-supervised learning model comprises the following steps: first step: acquiring large-scale unlabeled pure audio data by a microphone device, adding a random reverberation interference term to the pure audio data by a room impulse response function, and linearly adding a noise interference term to obtain a large-scale unlabeled noise and reverberation-containing audio database; clustering and labeling the large-scale unlabeled pure audio data; and inputting the large-scale unlabeled noise and reverberation-containing audio database into a self-supervised learning model, taking a clustering result of corresponding large-scale unlabeled pure audio data as a pseudo label, and taking the pseudo label as a learning target to complete a training task of the self-supervised learning model; second step: constructing a speaker recognition network model by the trained self-supervised learning model in the first step, and training and synchronously verifying the speaker recognition network model; and third step: extracting testing audio data from the large-scale unlabeled noise and reverberation-containing audio database, and making a corresponding speaker classification label; and testing a brand new speech signal by the speaker recognition network model obtained in the second step. Further, in the first step, a method for clustering and labeling comprises: obtaining a Mel-frequency cepstral coefficient of the large-scale unlabeled pure audio data, and clustering and labeling the large-scale unlabeled pure audio data by a K-means++ algorithm. Further, in the second step, the speaker recognition network model comprises a front end and a rear end, wherein the trained self-supervised learning model in the first step is migrated to the front end as a feature extractor of speech data, and the rear end sequentially comprises several one-dimensional convolutional networks, time delay neural networks, aggregation and pooling networks and full connection output networks. Further, in the second step, training audio data are extracted from the large-scale unlabeled noise and reverberation-containing audio database to make the corresponding speaker classification label, the labeled training audio data are input 2 into the constructed speaker recognition network model, and a speaker classification serial number is taken as a label to complete the training and synchronous verification of the speaker recognition network model. Further, in the second step, the extracted training audio data account for 5%-10% of the large-scale unlabeled noise and reverberation-containing audio database. Further, in the third step, audio data excluding the training audio data extracted from the large-scale unlabeled noise and reverberation-containing audio database are taken as the testing audio data. Further, in the second step, the one-dimensional convolutional network and the time delay neural network both contain a linear rectification function and a batch regularization function. Further, in the second step, a process of training and synchronous verification specifically comprises: transforming audio features output by a front-end Transformer encoder into speech features [B, D, T] through one-dimensional convolution; inputting the speech features [B, D, T] into two continuous one-dimensional convolutional neural networks to obtain speech features [B, F, T]; inputting the speech features [B, F, T] into three continuous time delay neural networks, wherein each time delay neural network contains a residual network unit and a squeeze-and-excitation unit; and allowing output results of the three continuous time delay neural networks to pass through a full connection layer, a linear rectification layer, a full connection layer and an S-type rectification layer in sequence, readjusting the operation, and outputting final results. Further, in the third step, the testing further comprises: calculating evaluation indexes and other error rates of calculated classification vectors, wherein associated evaluation indexes are an error acceptance rate, an error rejection rate and thresholds, different thresholds are set to correspond to error acceptance rates and error rejection rates of the different thresholds, and an equal error rate refers to a case that the error acceptance rate is equal to the error rejection rate. Further, a calculation method for the equal error rate comprises: 1) inputting the speech features output by the self-supervised learning model into a speaker recognition system to obtain the classification vectors; 2) calculating a cosine similarity between the classification vectors corresponding to various audio data, and defining the cosine similarity as a target score; 3) adding a target label and the target score into a scoring list; and 4) gradually increasing the threshold from 0 to 0.99 at an increase rate of 0.01 to obtain corresponding error acceptance rate and error rejection rate, and then calculating an equal error rate of speech data of a whole test set, wherein the lower the equal error rate is, the better the recognition effect is. Compared with the prior art, the speaker recognition method in the real sound field environment based on the self-supervised learning model provided by the present invention has the following beneficial effects. A lot of pure speech data from different speakers are processed by random reverberation to produce a reverberation-containing speech signal simulating a real conference room, then the reverberation-containing speech signal is linearly superimposed with a specially made noise signal at a very low signal-to-noise ratio, and a produced reverberation and noise-containing speech signal is close to a speech signal of the real conference room to the greatest extent. Noise and reverberation-containing mixed speech is taken as an input, and the pseudo label made from a large number of unlabeled pure signals is taken as an iteration target. The self-supervised learning model enhances the unlabeled data in three aspects: noise reduction, reverberation reduction and late-stage model random interference reduction, and the pseudo label provides a dynamic adaptive filtering extraction unit for the subsequent speaker recognition method to some extent, thereby improving the robustness of the features to strong noise and reverberation. A cross-channel fusion convolution mechanism is introduced into the residual network, which enhances context attention and fusion of adjacent channel areas in the speech features; a non-average attention mechanism is introduced into a squeeze-and-excitation network, so that each channel averages weight information at a time-domain frame level, thereby enhancing a contribution difference of different frames to speaker information; and a multi-head attentive statistics pooling mechanism is introduced to replace an attention score statistics pooling operation, which enhances a contribution difference of different frequency bands to a time-domain frame, and divides a whole channel space into several independent sub-channel spaces, thereby reducing mutual interference between the sub-channel spaces as much as possible. The speaker recognition network is optimized from many 4 perspectives, and according to characteristics of noise and reverberation in the feature extractor, the residual network, the squeeze-and-excitation network and the pooling layer network are improved to better deal with severely damaged speeches. BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a flow chart of a speaker recognition method in a real sound field environment based on a self-supervised learning model; FIG. 2 is a flow chart of training of the self-supervised learning model; and FIG. 3 is a flow chart of training of a speaker recognition network model. DETAILED DESCRIPTION The present invention is further explained and described hereinafter with reference to the drawings and specific embodiments. As shown in FIG 1, a speaker recognition method in a real sound field environment based on a self-supervised learning model comprises the following steps. In first step, large-scale unlabeled audio data are acquired in a low-noise and low-reverberation sound field by a microphone device, and the audio data contain almost no typical noise and reverberation interference term, so that the audio data may be considered as pure audio data. A random reverberation interference term is added to the pure audio data by a room impulse response function, and a noise interference term is linearly added to obtain a large-scale unlabeled strong noise and reverberation-containing audio database. A Mel-frequency cepstral coefficient of the large-scale unlabeled pure audio data is obtained, and the large-scale unlabeled pure audio data are clustered and labeled by a K-means++ algorithm. The large-scale unlabeled noise and reverberation-containing audio database is input into the self-supervised learning model, a clustering result of MFCC kmeans++ of corresponding large-scale unlabeled pure audio data is taken as a pseudo label, and the pseudo label is taken as a learning target to complete a training task of the self-supervised learning model, wherein the self-supervised learning model comprises a convolutional neural network encoder, a Transformer encoder and a projection matrix in sequence. In this embodiment, large-scale unlabeled speech data containing a strong noise and reverberation interference item are taken as input data, and an MFCC clustering 5 result of corresponding pure speech data is taken as label data. As shown in FIG. 2, a speech x with a length L is transformed into T (T=L / 320) features (a number of channels is 512) after seven convolution operations by the convolutional neural network encoder, wherein step sizes of the seven convolution operations are [5, 2, 2, 2, 2, 2, 2] and sizes of seven convolution kernels are [10, 3, 3, 3, 3, 2, 2], The convolutional neural network encoder participates in parameter adjustment in training. yk=bk+^^Wvk-xv (1) wherein, n, m, k and b refer to a convolution kernel size, an input channel, an x W v output channel and an output channel offset respectively, y, yk and Jk refer to an input speech component, a convolution parameter and an output speech component respectively, and T features are divided into two categories: one category participates in a mask (an initial appearance position of the mask is randomly selected, and 10 consecutive features are masked, with an assigned value 0) and the other category does not participate in the mask, wherein the features participating in the mask account for 20% of the total features. After the features participating in the mask and the features not participating in the mask pass through the Transformer encoder, , , p , m’ 5 ^T, —m j um ’ ’ —um , • > output feature sequences are L 1 J and L 2 J respectively. The Transformer encoder participates in parameter adjustment in training. The output feature sequences are transformed into corresponding probability density functions through the projection matrix and a corresponding codebook number of the pseudo label, and the corresponding probability density functions of the feature sequence participating in the mask and the feature sequence not participating in the mask are Formula (2) and Formula (3) respectively: , । ~ x exp (sim(A-ot_m,ec_tg} / Pf-m \cxp ^sim^A. exp(sim(A-ot_um,ec_tg} / A Pf-um (q x> d = ( ( V 7 exp [sim{A-ot_um,ec,) / T) wherein, t is a specific time-domain frame serial number, / f" \ I / is a probability density of the feature sequence participating in the mask, f~um' ' ' is a probability density of the feature sequence not participating in the mask, A is a backwardly propagable projection matrix, 0t is a certain frame in the feature sequence, ec_tg is a codebook after adding the pseudo label, c represents a certain specific classification selected by the pseudo label, ec, is each codebook after C classification, wherein each codebook participates in parameter adjustment in training, t is an adjusting scale parameter, and sim refers to calculating a cosine similarity between two tensors. A total cross entropy loss function is calculated for Formula (2) and Formula (3), as shown in Formula (4), L = M +(1-2)-1 m \ / um wherein, A is an adjustment factor, Lm is a loss function obtained by finally calculating values of the features participating in the mask, Lum is a loss function obtained by finally calculating values of the features not participating in the mask, and the loss functions are defined as Formula (5) and Formula (6), Lm =-Yt^ogPf_m(c\X,A v ’ (5) ^um — V ’ (6) wherein, lower comer marks m and um represent speech features participating in the mask and not participating in the mask respectively. The self-supervised learning model takes an MFCC clustering value of pure speech as the pseudo label for iterative learning by 20 rounds of learning. In an iterative process, a convolutional neural network coding layer, a Transformer coding layer and a projection matrix all participate in parameter learning and adjustment. When the self-supervised learning model is subsequently used as the feature extractor, the projection matrix is removed, and the convolutional neural network coding layer and the Transformer coding layer do not participate in parameter adjustment. Meanwhile, a loss function of a downstream task is used for parameter learning, and specifically, for a speaker recognition task, a normalized exponential function of an additive angular margin is used as a corresponding loss function. In second step, a speaker recognition network model is constructed, wherein the speaker recognition network model comprises a front end and a rear end, wherein the trained self-supervised learning model in the first step is migrated to the front end as a feature extractor of speech data, and the rear end sequentially comprises several one-dimensional convolutional networks, time delay neural networks, aggregation and pooling networks and full connection output networks. A small part of audio data (5%-10%) are extracted from the large-scale unlabeled noise and reverberation-containing audio database in the first step as training audio data to make the corresponding speaker classification label, the labeled training audio data are input into the constructed speaker recognition network model, and several speaker classification serial numbers are taken as labels to complete the training and synchronous verification of the speaker recognition network model. A specific method for training and verifying the speaker recognition model is as follows. A small part (5%-10%) of the large-scale unlabeled noise and reverberation-containing audio data in the first step is taken as an input, and a corresponding speaker serial number is a classification label. As shown in FIG. 3, an original dimension of a speech signal is [B, L], The trained self-supervised learning model in the first step is used as a speech feature extractor, wherein, in order to make the self-supervised learning model smoothly cascade with a rear-end speaker identification network, audio features output by the Transformer encoder are transformed into a general format of the speaker recognition network through one-dimensional convolution, that is, speech features [B, D, T], Each of the following neural networks is operated by a linear rectification function and a batch regularization parameter. The speech features [B, D, T] are input into two continuous one-dimensional convolutional neural networks to obtain speech features [B, F, T] for dimension adjustment, wherein convolution kernel sizes of the two one-dimensional convolutional neural networks are [1, 5] respectively, an input channel is 80, and an output channel is 256, The speech features [B, F, T] are input into three continuous time delay neural networks to achieve aggregation and fusion extraction of feature information, wherein each layer of time delay neural network module contains a fused residual network and a squeeze-and-excitation - attention average network. In addition, the one-dimensional convolution network in the time delay neural network adopts a dilated convolution technology to expand a receptive field, and convolution dilatation coefficients are [2, 3, 4] respectively, and each time delay neural network contains a residual network unit and a squeeze-and-excitation unit. The residual network unit is as shown in Formula (8), and an improved fused residual network unit is as shown in FIG. 3, which has a shape similar to a U-type network. An input x is divided into s parts, each part is expressed as xh and a number of feature channels in each group is 1 / s of a number of original channels; excluding x15 each group of features may be subjected to a 3*3 convolution, and this convolution operation is recorded as excluding x} and x2, an output of a previous group ) is added to an ith group of features xh and an addition result is subjected to the Ky ) operation; and in a descending stage of the U-type network, the ith feature is subjected to the 3*3 convolution and fused into an (i + l)th feature, in an ascending stage of the U-type network, the ith feature is subjected to the 3*3 convolution and fused into an (i — l)th feature, and in addition, an nth feature is convolved and directly sent to a next layer, then subjected to a 1*1 convolution, and fused together, so as to finally achieve context attention and fusion of adjacent channel areas. Features of different receptive fields and multiple scales can be extracted, so that global features and local features are effectively extracted, (8) The improved squeeze-and-excitation unit is as shown in FIG. 3, referring to Formula (9) to Formula (13), wherein vc represents a cth convolution kernel, and x^ represents an ith input. Extracted features are obtained through a transformation operation by Formula (9), wherein a dimension is H*W*C, H and W are a height and a width respectively, C is a number of channels, is a specific feature component of X, and Vc is a specific convolution parameter. uc = vc * X = X vc * Xs 5=1 (9) Formula (10) and Formula (11) represent a non-average squeeze operation, which is used for calculating distribution of values of various layers of channels, and transforming H*W*C into an output of 1*1*C. Specifically, x(B,F,T) is subjected to compression, restoration and randomization in sequence to obtain a weight coefficient a for all F channels and each T channel, and a dimension is [B, F, 1], A mean of x is obtained by a in a T dimension. A corresponding linear operation of and b3 represents compression, a corresponding linear operation of V2 and b4 represents restoration to an original state, then the weight coefficient a is obtained through a normalized exponential operation, and then a Hadamard product between the weight coefficient and x is solved to obtain zc. a = Softmax(V2 ■ tanh (V^-x + b^+b^ zc=lLat-xt (10) (11) Formula (12) represents an excitation operation, multiplication of by z is a full connection layer operation, a dimension of 1¾ is C / r * C, and r is a scaling parameter, which is intended to reduce a number of channels, thereby reducing an amount of calculation; a dimension of z is 1*1*C, and a result of W}z is l*l*C / r; then, the result passes through a linear rectification layer, and a dimension of an output remains unchanged; then, multiplication of the output by W2 is also a full connection layer operation, a dimension of W2 is C*C / r, and a dimension of an output is 1*1 *C; and finally, s is obtained through an S-type rectification function, 5 = ^ex(z^) = a(g(z^)) = (12) Formula (13) represents a channel-by-channel adjustment operation, and a Hadamard product between an adjustment coefficient of each channel and original X feature information is solved to obtain an adjusted output c, x = F , (u ,s }= s u c scale \ c >c J c c (13) Specifically, the output [B, F, 1] obtained by Formula (11) passes through the full connection layer, the linear rectification layer, the full connection layer and the S-type rectification layer in sequence, and the operation is readjusted. A corresponding linear operation of 1¾ and b} represents squeeze to compress the F channel, and a corresponding linear operation of W2 and b2 represents excitation to restore the F channel to an original state; and Xf is expanded to Xft (T dimension), and then Xft is readjusted with x to obtain x. xf = sigmoid ■ relu ■ zc + b^ + b^ (14) Outputs of the previous three time delay neural networks, which have a dimension (B, F, T) respectively, are connected in series according to the feature dimension, so that a dimension of a feature h after series connection is (B, 3*F, T). A mean and a standard deviation are calculated in the T dimension, which have a dimension (B, 3*F) respectively, and the mean and the standard deviation are stacked in the T dimension for T times to restore to (B, 3*F, T); and the feature, the mean and the standard deviation are spliced in an F dimension to obtain a feature x with a dimension (B, 9*F, T). Specifically, the aggregated x is divided into k non-overlapping sub-spaces along the channel dimension F, xt = [x*,x?, ...,xk] with (1 <t <T, xlt G Rd^k). Then, features xl = [x{,X2, ...,x^](l <i <k) of the same sub-channel space in different time dimensions are extracted, and a factor el of each sub-channel space is calculated by Formula (16). ™ F •tanh (JF 4- / f ) + q wherein, Wl G RFxWk^ is a first linear unit, vl G RFxF is a second linear unit, and bl G RFxT and cl G RFxT are two offset variables. Channel isolation self-attention scoring is used herein. An exponential normalization function is applied to each el in the time dimension T to obtain a percentage a1 of data frames for different sub-channel spaces, which reflects importance distribution of data frames at different times. (17) k a1 are spliced according to the channel dimension to obtain a (the weight coefficient). At this time, a sum of each line of features in a feature map in the T dimension is equal to 1, which is an effect of activation of the normalized exponential function; and because of having the same dimension as h, a may be regarded as an attention score, a mean and a standard deviation based on attention are calculated by Formula (18) and Formula (19), obtained features all have the dimension (B, 3*F), and then the features are connected in series according to the feature dimension to obtain a final output of an attentive statistics pooling layer, which has a dimension (B, 6*F). T / =l (18) &“JE«A 0¾-po^ (19) wherein, a‘ , and &represent a mean, an attention score corresponding to each frame, a specific feature corresponding to each frame, and a standard deviation respectively, represents a Hadamard product operation, features with the dimension (B, 6*F) are sent to the full connection layer to obtain a vector (B, 192), and the vector is used for calculating a loss function of an additive angular margin to finally obtain a classification vector. In third step, excluding the training audio data extracted in the second step, testing audio data are extracted from the large-scale unlabeled noise and reverberation-containing audio database, and a corresponding speaker classification label is made; and a brand new speech signal is tested by the speaker recognition network model obtained in the second step, wherein the brand new speech signal is a speech signal that has never appeared in a training set of the audio database. Evaluation indexes and other error rates of calculated classification vectors are calculated, wherein associated evaluation indexes are an error acceptance rate, an error rejection rate and thresholds, different thresholds are set to correspond to error acceptance rates and error rejection rates of the different thresholds, and an equal error rate refers to a case that the error acceptance rate is equal to the error rejection rate. A calculation method for the equal error rate comprises: 1) inputting the speech features output by the self-supervised learning model into a speaker recognition system to obtain the classification vectors; 2) calculating a cosine similarity between the classification vectors corresponding to various audio data, and defining the cosine similarity as a target score; 3) adding a target label and the target score into a scoring list; and 4) gradually increasing the threshold from 0 to 0.99 at an increase rate of 0.01 to obtain corresponding error acceptance rate and error rejection rate, and then calculating an equal error rate of speech data of a whole test set, wherein the lower the equal error rate is, the better the recognition effect is. An effect of this embodiment is verified by the following comparative experiment. Pure speaker speeches used in the experiment come from an open-source Chinese speech corpus, zhvoice (abbreviated as ZH). The ZH corpus is composed of 8 open-source data sets, such as Zhaidatatang and Zhaishell, after noise reduction and mute removal, and contains about 3,200 speakers and 900 hours of audio data at a sampling frequency of 16 kHz. The ZH is expanded to a version with noise and reverberation for a conference room scene, which is called NR-ZH for short. For the NR-ZH, it is necessary to simulate and produce a speech with reverberation first, and a room impulse response is produced through a pyroom acoustics library. It is worth noting that an expanded version of T60 is used in the NR-ZH considering that a conference room usually has a large space. In addition, pyroom acoustics uses dual sound sources and dual microphones by default, while the simplest speaker recognition system only needs a monophonic speech. Two groups of identical sound sources and microphone parameters are input into a whamroom module, and only a left channel of a first microphone is used as a reverberation output signal. Then, a noise signal is randomly selected for each speech signal with reverberation, and linear superposition is completed by a random signal-to-noise ratio ([0, 0.5]dB). If a noise length is less than a speech length, a tail part of noise is filled with a front part of noise to make the noise length equal to the speech length. The noise comes from acoustic signals (in a total of 12 categories, comprising air conditioners, fans, vibration of mobile phones, coughing, sneezing, walking, page turning of files, keyboards, and the like, 1950 items in total) obtained by paid download from Soundsnap website and acoustic signals (sampled at 16 khz, with a length cut into 3 seconds, 480 items in total) actually recorded in a conference room during working by a U-PHORIA UMC202HD acoustic sensor produced by behringer. A baseline model comprises TDNN (512 channels), Resnet-SE (each residual module is set as [32, 64, 128, 256] channels), Res2net (each residual module is set as [64, 128, 256, 512] channels), Eres2net (each residual module is set as [32, 64, 128, 256] channels), ECAPA-TDNN (512 channels) and CAM++ (32 channels). By comparing a plurality of baseline models, an index superiority of the speaker recognition method provided by the present invention may be verified. In the real sound field environment, because there are other speakers and noise, spectrograms of 13 sound, reverberation and noise are difficult to be distinguished. After processing by the self-supervised learning model + the speaker recognition network, the best recognition effect is achieved. Those described above are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made without departing from the spirit and principle of the present invention shall fall within the scope of protection of the present invention.

Claims

1. A speaker recognition method in a real sound field environment based on a self-supervised learning model, comprising the following steps:first step: acquiring large-scale unlabeled pure audio data by a microphone device, adding a random reverberation interference term to the pure audio data by a room impulse response function, and linearly adding a noise interference term to obtain a large-scale unlabeled noise and reverberation-containing audio database; clustering and labeling the large-scale unlabeled pure audio data; andinputting the large-scale unlabeled noise and reverberation-containing audio database into a self-supervised learning model, taking a clustering result of corresponding large-scale unlabeled pure audio data as a pseudo label, and taking the pseudo label as a learning target to complete a training task of the self-supervised learning model;second step: constructing a speaker recognition network model by the trained self-supervised learning model in the first step, and training and synchronously verifying the speaker recognition network model; andthird step: extracting testing audio data from the large-scale unlabeled noise and reverberation-containing audio database, and making a corresponding speaker classification label; and testing a brand new speech signal by the speaker recognition network model obtained in the second step.

2. The speaker recognition method in the real sound field environment based on the self-supervised learning model according to claim 1, wherein, in the first step, a method for clustering and labeling comprises: obtaining a Mel-frequency cepstral coefficient of the large-scale unlabeled pure audio data, and clustering and labeling the large-scale unlabeled pure audio data by a K-means++ algorithm.3 The speaker recognition method in the real sound field environment based on the self-supervised learning model according to claim 1 or 2, wherein, in the second step, the speaker recognition network model comprises a front end and a rear end, wherein the trained self-supervised learning model in the first step is migrated to the front end as a feature extractor of speech data, and the rear end sequentially comprises several one-dimensional convolutional networks, time delay neural networks, aggregation and pooling networks and full connection output networks.

4. The speaker recognition method in the real sound field environment based on the self-supervised learning model according to claim 1 or 2, wherein, in the second step, training audio data are extracted from the large-scale unlabeled noise and reverberation-containing audio database to make the corresponding speaker classification label, the labeled training audio data are input into the constructed speaker recognition network model, and a speaker classification serial number is taken as a label to complete the training and synchronous verification of the speaker recognition network model.

5. The speaker recognition method in the real sound field environment based on the self-supervised learning model according to claim 4, wherein, in the second step, the extracted training audio data account for 5%-10% of the large-scale unlabeled noise and reverberation-containing audio database.

6. The speaker recognition method in the real sound field environment based on the self-supervised learning model according to claim 4, wherein, in the third step, audio data excluding the training audio data extracted from the large-scale unlabeled noise and reverberation-containing audio database are taken as the testing audio data.

7. The speaker recognition method in the real sound field environment based on the self-supervised learning model according to claim 1 or 2, wherein, in the second step, the one-dimensional convolutional network and the time delay neural network both contain a linear rectification function and a batch regularization function.

8. The speaker recognition method in the real sound field environment based on the self-supervised learning model according to claim 1 or 2, wherein, in the second step, a process of training and synchronous verification specifically comprises:transforming audio features output by a front-end Transformer encoder into speech features [B, D, T] through one-dimensional convolution;inputting the speech features [B, D, T] into two continuous one-dimensional convolutional neural networks to obtain speech features [B, F, T];inputting the speech features [B, F, T] into three continuous time delay neural networks, wherein each time delay neural network contains a residual network unit and a squeeze-and-excitation unit; andallowing output results of the three continuous time delay neural networks to pass through a full connection layer, a linear rectification layer, a full connection layer and an S-type rectification layer in sequence, readjusting the operation, andoutputting final results.

9. The speaker recognition method in the real sound field environment based on the self-supervised learning model according to claim 1 or 2, wherein, in the third step, the testing further comprises: calculating evaluation indexes and other error rates of calculated classification vectors, wherein associated evaluation indexes are an error acceptance rate, an error rejection rate and thresholds, different thresholds are set to correspond to error acceptance rates and error rejection rates of the different thresholds, and an equal error rate refers to a case that the error acceptance rate is equal to the error rejection rate.

10. The speaker recognition method in the real sound field environment based on the self-supervised learning model according to claim 9, wherein, a calculation method for the equal error rate comprises:1) inputting the speech features output by the self-supervised learning model into a speaker recognition system to obtain the classification vectors;2) calculating a cosine similarity between the classification vectors corresponding to various audio data, and defining the cosine similarity as a target score;3) adding a target label and the target score into a scoring list; and4) gradually increasing the threshold from 0 to 0.99 at an increase rate of 0.01 to obtain corresponding error acceptance rate and error rejection rate, and then calculating an equal error rate of speech data of a whole test set, wherein the lower the equal error rate is, the better the recognition effect is.T +44(0)30 0300 2000A

Images