Voiceprint recognition model training method and system based on big data self-supervision
Through the big data self-supervision method, the distributed host performs recording data cleaning and slicing processing, and conducts federated learning training on the server, solving the problems of low training efficiency and insufficient diversity of traditional voiceprint recognition models, and realizing the massive data model Improved training and recognition accuracy.
Patent Information
- Application Number
- CN202210986290.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-08-17
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2042-08-17
Smart Images

Figure CN115472167B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of voiceprint recognition, and in particular to a voiceprint recognition model training method and system based on big data self-supervision. Background Art
[0002] Voiceprint recognition is a type of biometric recognition technology, also known as speaker recognition, which includes speaker identification and speaker confirmation. Voiceprint recognition is to convert sound signals into electrical signals and then use computers to recognize them. The traditional voiceprint recognition method is to input voiceprint data into a voiceprint recognition model, which is then used to recognize the voiceprint and obtain the recognition result. The voiceprint recognition model is learned from training data through machine learning. The training data can be various recording data, which can include annotated recording data; the recording data can also include unannotated recording data, such as closed-circuit television recording data, outdoor interview data, etc.
[0003] However, although the traditional voiceprint recognition model has high training efficiency, its richness is low and its diversity is not as large as massive data, which is not conducive to model generalization, and the amount of data may not be as large as that of models with large parameters. Summary of the invention
[0004] Based on this, in order to solve the above technical problems, a voiceprint recognition model training method and system based on big data self-supervision are provided, which can realize massive data model training.
[0005] A voiceprint recognition model training method based on big data self-supervision, the method comprising:
[0006] Each distributed host obtains multi-source recording data, distributes and stores the recording data from different sources, and cleans the recording data to obtain labeled recording data and unlabeled recording data;
[0007] Each of the distributed hosts slices the labeled recording data and the unlabeled recording data, adds noise interference to the sliced recording data, obtains sample pairs, and inputs the sample pairs into a self-supervised learning model, performs federated learning model training with the server, and obtains a voiceprint recognition basic model through training on the server;
[0008] The server performs supervised learning and fine-tuning on the model parameters of the basic model based on the labeled recording data to obtain a voiceprint recognition fine-tuning model.
[0009] In one embodiment, each of the distributed hosts cleans the recording data to obtain the recorded data with labels and the recorded data without labels, including:
[0010] Each of the distributed hosts uses a resampling algorithm to adjust the sampling frequency of the recording data to a target sampling frequency;
[0011] Each of the distributed hosts obtains a signal-to-noise ratio corresponding to the recording data and obtains a target signal-to-noise ratio; and deletes the recording data having a signal-to-noise ratio less than the target signal-to-noise ratio;
[0012] Each of the distributed hosts performs non-voice removal processing on the remaining recorded data to obtain recorded data with labels and recorded data without labels.
[0013] In one embodiment, each of the distributed hosts performs slicing processing on the labeled recording data and the unlabeled recording data to obtain sample pairs, including:
[0014] Each of the distributed hosts obtains a slicing condition, and determines the recording data to be sliced from the tagged recording data and the untagged recording data according to the slicing condition;
[0015] Each of the distributed hosts randomly cuts the recording data to be sliced to obtain slices, and each of the slices does not overlap in time;
[0016] Each of the distributed hosts forms a sample pair with each of the slices.
[0017] In one embodiment, each of the distributed hosts forms a sample pair with each of the slices, including:
[0018] Each of the distributed hosts forms a positive sample pair from each of the slices obtained by cutting the same recording sample;
[0019] Each of the distributed hosts forms a negative sample pair with each of the slices obtained by cutting from different recording samples.
[0020] In one embodiment, each of the distributed hosts adds noise interference to the sliced recording data, including:
[0021] Each of the distributed hosts adds additive noise interference and convolution noise interference to each of the slices respectively.
[0022] In one embodiment, the sample pair is input into a self-supervised learning model, and federated learning model training is performed with the server, and a voiceprint recognition basic model is obtained by training on the server, including:
[0023] The server sends the encrypted initial model parameters to each of the distributed hosts;
[0024] Each of the distributed hosts inputs the sample pairs into the self-supervised learning model for model training, obtains the initial model training gradient, and transmits it to the server in encrypted form;
[0025] The server receives the encrypted initial model training gradient and decrypts it, and the server adjusts the initial model parameters according to the initial model training gradient to obtain the target model parameters and encrypts them;
[0026] The server sends the encrypted target model parameters to each of the distributed hosts, and receives the encrypted target model training gradients fed back by each of the distributed hosts based on the target model parameters. The server decrypts the target model training gradients. When the target model training gradients converge to 0, the server obtains the voiceprint recognition basic model based on the target model parameters training.
[0027] In one embodiment, the server receives and decrypts the encrypted initial model training gradient, and the server adjusts the initial model parameters according to the initial model training gradient, including:
[0028] The server receives the encrypted initial model training gradient and performs a decryption operation, and performs a weighted sum calculation on the initial model training gradient returned by each distributed host according to the credibility weight. The server adjusts the initial model parameters according to the weighted summed gradient.
[0029] A voiceprint recognition model training system based on big data self-supervision, the system comprising:
[0030] Each distributed host is used to obtain multi-source recording data, distribute and store the recording data from different sources, and clean the recording data to obtain labeled recording data and unlabeled recording data;
[0031] Each of the distributed hosts is further used to slice the labeled recording data and the unlabeled recording data, add noise interference to the sliced recording data to obtain sample pairs, and input the sample pairs into the self-supervised learning model to perform federated learning model training with the server, and train on the server to obtain a basic model for voiceprint recognition;
[0032] The server is used to perform supervised learning and fine-tune the model parameters of the basic model based on the labeled recording data to obtain a voiceprint recognition fine-tuning model.
[0033] The above-mentioned method and system for training a voiceprint recognition model based on big data self-supervision obtain multi-source recording data through each distributed host, distribute the recording data from different sources, and clean the recording data to obtain labeled recording data and unlabeled recording data; each of the distributed hosts slices the labeled recording data and the unlabeled recording data, respectively adds noise interference to the sliced recording data to obtain sample pairs, and inputs the sample pairs into the self-supervised learning model, performs federated learning model training with the server, and obtains a voiceprint recognition basic model through training on the server; the server performs supervised learning fine-tuning on the model parameters of the basic model based on the labeled recording data to obtain a voiceprint recognition fine-tuning model. After collecting multi-source recording data and cleaning the data, the unlabeled recording data is used for self-supervised training, and the labeled data is used for model parameter fine-tuning, which can improve the model training efficiency; and the multi-source data stored on different distributed hosts by the server are coordinated by horizontal federated learning for large-scale training of voiceprint recognition models, which can achieve massive data model training. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 This is a diagram of an application environment of a voiceprint recognition model training method based on big data self-supervision in one embodiment;
[0035] Figure 2 1 is a flow chart of a method for training a voiceprint recognition model based on big data self-supervision in one embodiment;
[0036] Figure 3 The figure is a schematic diagram of data interaction between the master host and each distributed host in one embodiment. DETAILED DESCRIPTION
[0037] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0038] The voiceprint recognition model training method based on big data self-supervision provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Figure 1As shown, the application environment includes a server 110 and various distributed hosts 120, wherein the server 110 can communicate with each distributed host 120. Each distributed host 120 obtains multi-source recording data, distributes and stores the recording data from different sources, and cleans the recording data to obtain labeled recording data and unlabeled recording data; each distributed host 120 slices the labeled recording data and the unlabeled recording data, respectively adds noise interference to the sliced recording data to obtain sample pairs, and inputs the sample pairs into the self-supervised learning model, performs federated learning model training with the server 110, and obtains the voiceprint recognition basic model through training on the server 110; the server 110 performs supervised learning fine-tuning on the model parameters of the basic model based on the labeled recording data to obtain the voiceprint recognition fine-tuning model.
[0039] In one embodiment, Figure 2 As shown, a voiceprint recognition model training method based on big data self-supervision is provided, comprising the following steps:
[0040] In step 202, each distributed host obtains multi-source recording data, stores the recording data from different sources in a distributed manner, and cleans the recording data to obtain labeled recording data and unlabeled recording data.
[0041] Each distributed host can obtain multi-source recording data, that is, recording data from various sources. For example, each distributed host can obtain multi-source voice data collected by customer service centers, telecom operators and other data providers through their own channels and machines, including telephone recordings, surveillance recordings, network video tracks, etc. Recording data from different sources can be distributed and stored in the data center on the collection side to protect privacy and security.
[0042] Each distributed host can perform data cleaning and preprocessing operations on the acquired multi-source recording data to obtain labeled recording data and unlabeled recording data. Among them, the cleaned labeled recording data and unlabeled recording data can be manually sampled in a certain proportion to ensure the voice quality. If the sample fails, the strictness of data cleaning will be increased. The cleaned unlabeled recording data can be used for subsequent large-scale self-supervised learning.
[0043] In step 204, each distributed host performs slice processing on the labeled recording data and the unlabeled recording data, adds noise interference to the sliced recording data, obtains sample pairs, and inputs the sample pairs into the self-supervised learning model, performs federated learning model training with the server, and obtains the basic model of voiceprint recognition by training on the server.
[0044] Each distributed host can slice the labeled recording data and the unlabeled recording data, that is, cut the unlabeled recording data that meets the slicing requirements to obtain slices, thereby obtaining sample pairs. The sample pairs can be input into the self-supervised learning model and trained with the server for the federated learning model, thereby training the basic model of voiceprint recognition on the server.
[0045] Each distributed host can add noise interference to the sliced recording data to achieve different channel data enhancements.
[0046] In step 206, the server performs supervised learning and fine-tuning on the model parameters of the basic model based on the labeled recording data to obtain a fine-tuned voiceprint recognition model.
[0047] In this embodiment, multi-source recording data is obtained through each distributed host, the recording data from different sources are distributedly stored, and the recording data is cleaned to obtain labeled recording data and unlabeled recording data; each distributed host slices the labeled recording data and unlabeled recording data, adds noise interference to the sliced recording data, obtains sample pairs, and inputs the sample pairs into the self-supervised learning model, performs federated learning model training with the server, and obtains the voiceprint recognition basic model through training on the server; the server fine-tunes the model parameters of the basic model based on the labeled recording data through supervised learning to obtain the voiceprint recognition fine-tuning model. After collecting multi-source recording data and cleaning the data, the unlabeled recording data is used for self-supervised training, which can improve the model training efficiency; and the multi-source data stored on different distributed hosts by the server are coordinated by horizontal federated learning for large-scale training of voiceprint recognition models, which can realize massive data model training.
[0048] In one embodiment, a voiceprint recognition model training method based on big data self-supervision is provided, which may also include a data cleaning process, wherein each distributed host uses a resampling algorithm to adjust the sampling frequency of the recording data to a target sampling frequency; each distributed host obtains the signal-to-noise ratio corresponding to the recording data, and obtains the target signal-to-noise ratio; the recording data with a signal-to-noise ratio less than the target signal-to-noise ratio is deleted; each distributed host performs non-voice removal processing on the remaining recording data to obtain labeled recording data and unlabeled recording data.
[0049] Each distributed host can use a resampling algorithm to adjust the sampling frequency of the recording data. The target sampling frequency can be 8KHz, and each distributed host can uniformly resample the multi-source recording data to 8KHz. Among them, the target signal-to-noise ratio can be -5dB, and each distributed host can discard recordings with a signal-to-noise ratio below -5dB, and remove non-speech invalid components in the recording through active sound detection based on spectral energy, thereby obtaining a recording data set.
[0050] In this embodiment, by resampling the collected multi-source recording data, deleting it according to the signal-to-noise ratio, and removing non-speech invalid components, the quality of the cleaned recording data can be improved.
[0051] In one embodiment, a voiceprint recognition model training method based on big data self-supervision is provided, which may also include a process of slicing unlabeled recording data, the specific process including: each distributed host obtains a slicing condition, and according to the slicing condition, determines the recording data to be sliced from the labeled recording data and the unlabeled recording data; randomly cuts the recording data to be sliced to obtain individual slices, and the slices do not overlap in time; and constitutes each slice into a sample pair.
[0052] The slicing condition may be pre-set, and each distributed host may determine the recording data to be sliced from the tagged recording data and the untagged recording data according to the slicing condition. For example, the slicing condition may be a voice sample with a duration of more than 8 seconds, and each distributed host may select a voice sample with a duration of more than 8 seconds from the tagged recording data and the untagged recording data for slicing.
[0053] Specifically, each distributed host may randomly cut the voice sample into K slices, and ensure that the K slices do not overlap in time, so that each slice constitutes a sample pair.
[0054] In this embodiment, the speech sample is randomly cut into slices to form sample pairs, which can be used for further model self-supervision initialization.
[0055] In one embodiment, a voiceprint recognition model training method based on big data self-supervision is provided, which may also include a process of forming sample pairs, and the specific process includes: each distributed host forms a positive sample pair with each slice obtained from the same recording sample; and each slice obtained from different recording samples forms a negative sample pair.
[0056] Each distributed host can form a positive sample pair (a, p) with each slice cut from the same recording sample, and a negative sample pair (a, n) with slices cut from different voice samples. In this embodiment, each distributed host can train a neural network model based on the positive sample pair (a, p) and the negative sample pair (a, n) by triple loss and gradient descent method: L1 = max(0, d(a, p) - d(a, n) + m), where d represents the distance metric and m represents the adjustable spacing parameter. The basic model can be obtained by such a training method.
[0057] In one embodiment, a voiceprint recognition model training method based on big data self-supervision is provided, which may also include a process of adding noise interference, and the specific process includes: each distributed host adds additive noise interference and curvature noise interference to each slice respectively.
[0058] Each distributed host can add additive noise interference and curvature noise interference to each slice. Specifically, for the slices {1, 2, ..., K} obtained by intercepting the same voice sample, different data enhancement methods are adopted for each slice K, that is, different channel data enhancements are achieved by adding additive noise interference nk and different curvature noise interference rk to the slices. For different slices k, the additive noise data and curvature noise data used are different. Among them, the additive noise data can be collected through channels such as telephone recording, field collection, and network collection. It belongs to environmental noise and call background noise in multiple scenarios (such as office, outdoor, and market). The curvature noise is obtained through room impulse response numerical simulation and indoor field collection under clean sound conditions.
[0059] In this embodiment, considering the consistency of channel conditions of different slices of the same speech sample, the speaker identity information and channel information of the sliced speech are decoupled by channel invariance training of adding additive noise interference and convolution noise interference to each slice.
[0060] In one embodiment, a voiceprint recognition model training method based on big data self-supervision is provided, which may also include a process of training a federated learning model, and the specific process includes: the server sends encrypted initial model parameters to each distributed host; each distributed host inputs a sample pair into a self-supervised learning model for model training, obtains an initial model training gradient, and encrypts and transmits it to the server; the server receives the encrypted initial model training gradient and decrypts it, and the server adjusts the initial model parameters according to the initial model training gradient, obtains the target model parameters and encrypts them; the server sends the encrypted target model parameters to each distributed host, and receives the encrypted target model training gradient fed back by each distributed host according to the target model parameters, and the server decrypts the target model training gradient. When the target model training gradient converges to 0, the server obtains a basic model for voiceprint recognition according to the target model parameter training.
[0061] like Figure 3 As shown, the server is the master host, and each distributed host is a distributed slave host t, wherein the master host can be connected to each distributed slave host t to facilitate data communication. Each distributed slave host t can be used for data collection and data processing. The master host can distribute the encrypted initial model parameter W to each distributed host t, and the initial model parameter W sent by the master host can be expressed as D(W). After receiving the encrypted initial model parameter D(W), each distributed host t can further perform data processing operations.
[0062] Each distributed host can input the sample pair into the self-supervised learning model for model training, obtain the initial model training gradient, and encrypt and transmit it to the server. Specifically, the sample pair can be input into the self-supervised learning model for model training, thereby obtaining the initial model training gradient. Figure 3 As shown, the initial model training gradient can be obtained using E t To represent, each distributed host t can generate the initial model training gradient E t Encrypted and sent back to the server, the master host.
[0063] The server can adjust the initial model parameters according to the initial model training gradient to make the recognition model obtained by training more accurate. After adjusting the initial model parameters, the server can obtain the target model parameters, and the server can encrypt the target model parameters and transmit them to each distributed host.
[0064] In this embodiment, for the voiceprint recognition service, several linear layers and multi-classification layers can be superimposed on the basic model of voiceprint recognition, and the angular space extra margin Softmax function is used as the loss function: Where N is the total number of small batch training data samples read in each iteration, s and m are the adjustable scaling factor and adjustable angular space margin, respectively, and y i represents the classification label of sample i (expressed as a one-hot vector), and the initial model training gradient can be obtained as E t .
[0065] After the server sends the encrypted target model parameters to each distributed host, each distributed host can decrypt the encrypted target model parameters, perform model training, obtain the target model training gradient, and encrypt it again for transmission to the server.
[0066] The server can decrypt the target model training gradient until the target model training gradient converges to 0. When the target model training gradient converges to 0, it means that the recognition model trained by the target model parameters can achieve the expected accuracy. At this time, the server can train the voiceprint recognition basic model according to the target model parameters.
[0067] Specifically, the server, i.e., the master host, can determine whether the target model training parameters are locally optimal. If so, a voiceprint recognition model is generated based on the target model training parameters. If not, the server sends the updated target model parameters to each distributed host, and each distributed host updates the model training gradient and returns the updated model training gradient. The master host then calculates based on the updated model training gradient until the target model training parameters are locally optimal, thereby generating a voiceprint recognition model.
[0068] In one embodiment, a voiceprint recognition model training method based on big data self-supervision is provided, which may also include a process of calculating a loss value. The specific process includes: the server receives the encrypted initial model training gradient and performs a decryption operation, and performs a weighted sum calculation on the initial model training gradient returned by each distributed host according to the credibility weight, and the server adjusts the initial model parameters according to the weighted sum of the gradients.
[0069] like Figure 3 As shown, the encrypted initial model training gradient E t It can be expressed as E(W r ), where r = 1, 2, 3, ... is used to represent the number of each distributed host.
[0070] The server, i.e. the master host, receives the encrypted initial model training gradient E t And perform decryption operations, and train the initial model gradient E transmitted back by each distributed host according to the credibility weight pt t Perform weighted sum calculation to obtain the weighted summed gradient.
[0071] It should be understood that, although the various steps in the above-mentioned flow charts are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the above-mentioned various flow charts may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.
[0072] In one embodiment, Figure 1 As shown, a voiceprint recognition model training system based on big data self-supervision is provided, including: a server 110 and each distributed host 120, wherein:
[0073] Each distributed host 120 is used to obtain multi-source recording data, distribute and store the recording data from different sources, and clean the recording data to obtain the recording data with labels and the recording data without labels;
[0074] Each distributed host 120 is further used to slice the recorded data with labels and the recorded data without labels, add noise interference to the sliced recorded data to obtain sample pairs, and input the sample pairs into the self-supervised learning model to perform federated learning model training with the server 110, and train the basic model of voiceprint recognition on the server 110;
[0075] The server 110 is used to perform supervised learning and fine-tune the model parameters of the basic model based on the labeled recording data to obtain a voiceprint recognition fine-tuning model.
[0076] In one embodiment, each distributed host 120 is also used to adjust the sampling frequency of the recorded data to a target sampling frequency using a resampling algorithm; obtain the signal-to-noise ratio corresponding to the recorded data, and obtain the target signal-to-noise ratio; delete the recorded data with a signal-to-noise ratio less than the target signal-to-noise ratio; perform non-voice removal processing on the remaining recorded data to obtain labeled recorded data and unlabeled recorded data.
[0077] In one embodiment, each distributed host 120 is also used to obtain slicing conditions, and determine the recording data to be sliced from the labeled recording data and the unlabeled recording data according to the slicing conditions; randomly cut the recording data to be sliced to obtain individual slices, and the slices do not overlap in time; and form sample pairs with each slice.
[0078] In one embodiment, each distributed host 120 is further configured to form a positive sample pair with each slice obtained from cutting a same recording sample; and form a negative sample pair with each slice obtained from cutting different recording samples.
[0079] In one embodiment, each distributed host 120 is further configured to add additive noise interference and convolution noise interference to each slice respectively.
[0080] In one embodiment, the server 110 is further used to send encrypted initial model parameters to each distributed host 120; each distributed host 120 is further used to input sample pairs into the self-supervised learning model for model training, obtain initial model training gradients, and encrypt and transmit them to the server 110; the server 110 is further used to receive and decrypt the encrypted initial model training gradients, and adjust the initial model parameters according to the initial model training gradients to obtain and encrypt target model parameters; the server 110 is further used to send the encrypted target model parameters to each distributed host 120, and receive the encrypted target model training gradients fed back by each distributed host 120 according to the target model parameters. The server 110 is further used to decrypt the target model training gradients. When the target model training gradients converge to 0, the server 110 is further used to obtain a basic model for voiceprint recognition according to the target model parameter training.
[0081] In one embodiment, the server 110 is also used to receive the encrypted initial model training gradient and perform a decryption operation, and perform a weighted sum calculation on the initial model training gradient returned by each distributed host 120 according to the credibility weight, and adjust the initial model parameters according to the weighted summed gradient.
[0082] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0083] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.
Claims
1. A voiceprint recognition model training method based on big data self-supervision, characterized in that: The method comprises: Each distributed host obtains multi-source recording data, stores the recording data from different sources in a distributed manner, and cleans the recording data to obtain labeled recording data and unlabeled recording data; Each of the distributed hosts slices the labeled recording data and the unlabeled recording data, adds noise interference to the sliced recording data respectively, obtains sample pairs, and inputs the sample pairs into the self-supervised learning model, performs federated learning model training with the server, and obtains the voiceprint recognition basic model through training on the server, including: the server sends encrypted initial model parameters to each of the distributed hosts; each of the distributed hosts inputs the sample pairs into the self-supervised learning model for model training to obtain the initial model training gradient, and encrypts and transmits it to the server; the server receives the encrypted initial model training gradient and decrypts it, the server adjusts the initial model parameters according to the initial model training gradient, obtains the target model parameters and encrypts them; the server sends the encrypted target model parameters to each of the distributed hosts, and receives the encrypted target model training gradient fed back by each of the distributed hosts according to the target model parameters, the server decrypts the target model training gradient, and when the target model training gradient converges to 0, the server obtains the voiceprint recognition basic model through training according to the target model parameters; The server performs supervised learning and fine-tuning on the model parameters of the basic model based on the labeled recording data to obtain a voiceprint recognition fine-tuning model.
2. The method for training a voiceprint recognition model based on big data self-supervision according to claim 1, characterized in that: Each of the distributed hosts performs data cleaning on the recording data to obtain labeled recording data and unlabeled recording data, including: Each of the distributed hosts uses a resampling algorithm to adjust the sampling frequency of the recorded data to a target sampling frequency; Each of the distributed hosts obtains a signal-to-noise ratio corresponding to the recording data and obtains a target signal-to-noise ratio; and deletes recording data having a signal-to-noise ratio less than the target signal-to-noise ratio; Each of the distributed hosts performs non-voice removal processing on the remaining recorded data to obtain recorded data with labels and recorded data without labels.
3. The method for training a voiceprint recognition model based on big data self-supervision according to claim 1, characterized in that: Each of the distributed hosts slices the labeled recording data and the unlabeled recording data to obtain sample pairs, including: Each of the distributed hosts obtains a slicing condition, and determines the recording data to be sliced from the tagged recording data and the untagged recording data according to the slicing condition; Each of the distributed hosts randomly cuts the recording data to be sliced to obtain slices, and each of the slices does not overlap in time; Each of the distributed hosts forms a sample pair with each of the slices.
4. The method for training a voiceprint recognition model based on big data self-supervision according to claim 3 is characterized in that: Each of the distributed hosts forms a sample pair from each of the slices, including: Each of the distributed hosts forms a positive sample pair from each of the slices obtained by cutting the same recording sample; Each of the distributed hosts forms a negative sample pair with each of the slices obtained by cutting different recording samples.
5. The method for training a voiceprint recognition model based on big data self-supervision according to claim 1, characterized in that: Each of the distributed hosts adds noise interference to the sliced recording data, including: Each of the distributed hosts adds additive noise interference and convolutional noise interference to each of the slices respectively.
6. The method for training a voiceprint recognition model based on big data self-supervision according to claim 1, characterized in that: The server receives and decrypts the encrypted initial model training gradient, and adjusts the initial model parameters according to the initial model training gradient, including: The server receives the encrypted initial model training gradient and performs a decryption operation, and performs a weighted sum calculation on the initial model training gradient returned by each distributed host according to the credibility weight. The server adjusts the initial model parameters according to the weighted summed gradient.
7. A voiceprint recognition model training system based on big data self-supervision, characterized by: The system comprises: Each distributed host is used to obtain multi-source recording data, distribute and store the recording data from different sources, and clean the recording data to obtain labeled recording data and unlabeled recording data; Each of the distributed hosts is further configured to slice the labeled recording data and the unlabeled recording data, add noise interference to the sliced recording data, obtain sample pairs, input the sample pairs into the self-supervised learning model, perform federated learning model training with the server, and train on the server to obtain a basic voiceprint recognition model, including: the server sends encrypted initial model parameters to each of the distributed hosts; each of the distributed hosts inputs the sample pairs into the self-supervised learning model for model training to obtain an initial model training gradient, and encrypts and transmits it to the server; the server receives the encrypted initial model training gradient and decrypts it, the server adjusts the initial model parameters according to the initial model training gradient, obtains the target model parameters and encrypts them; the server sends the encrypted target model parameters to each of the distributed hosts, and receives the encrypted target model training gradient fed back by each of the distributed hosts according to the target model parameters, the server decrypts the target model training gradient, and when the target model training gradient converges to 0, the server trains according to the target model parameters to obtain a basic voiceprint recognition model; The server is used to perform supervised learning and fine-tune the model parameters of the basic model based on the labeled recording data to obtain a voiceprint recognition fine-tuning model.
8. The voiceprint recognition model training system based on big data self-supervision according to claim 7 is characterized in that: Each of the distributed hosts is further configured to adjust the sampling frequency of the recorded data to a target sampling frequency using a resampling algorithm; obtain a signal-to-noise ratio corresponding to the recorded data and a target signal-to-noise ratio; delete the recorded data having a signal-to-noise ratio less than the target signal-to-noise ratio; and perform non-speech removal processing on the remaining recorded data to obtain labeled recorded data and unlabeled recorded data.
9. The voiceprint recognition model training system based on big data self-supervision according to claim 7 is characterized in that: Each of the distributed hosts is further configured to obtain a slicing condition and, based on the slicing condition, determine the recording data to be sliced from the tagged recording data and the untagged recording data; The recorded data to be sliced are randomly cut to obtain slices, and the slices do not overlap in time; and the slices constitute sample pairs.
