A method and system for unregistered voiceprint recognition

Abstract

The application discloses a kind of free registration voiceprint identification methods, including obtaining to be measured voice data, the voiceprint feature of the voice data to be measured is extracted;The voiceprint feature of the voice data to be detected is compared with the voiceprint feature of the registered user in database register table, if no matching registered user is found, it is identified as new user and a new user ID is allocated for the new user, and the voiceprint feature of the user and new user ID are stored in register table and return new user ID;If matching registered user is found, the user ID of the registered user is returned.It also discloses a kind of free registration voiceprint identification system.The application can also carry out voiceprint identification for the user without prior registration, removes the cumbersome registration process in prior art voiceprint identification, solves the problem that prior art voiceprint identification needs the registration voice data of to be identified user, and cannot identify the user without entering registration voice data.

Description

Technical Field

[0001] This invention relates to the field of voiceprint recognition technology, specifically, to a registration-free voiceprint recognition method and system. Background Technology

[0002] With the development of deep learning technology, voiceprint recognition has made continuous breakthroughs based on deep speech processing technology, and its acceptance has grown significantly. This has also placed higher demands on voiceprint recognition systems, requiring not only high accuracy but also robustness in complex scenarios. In real-world applications, ideal recording studio environments are unattainable; received speech data often contains various reverberations and noise interference. Furthermore, the channel characteristics of speech data vary depending on the usage scenario and the device used. Traditional voiceprint recognition models, trained in clean entry environments with similar or limited channel conditions, exhibit poor robustness to far-field and noisy speech. Existing applications also require voiceprint recognition for unregistered users—the ability for the system to distinguish different speakers and record their information. However, traditional voiceprint recognition models fail to differentiate speakers when using unregistered speech as a reference. Therefore, existing models cannot meet current needs. Summary of the Invention

[0003] The purpose of this invention is to provide a registration-free voiceprint recognition method and system to solve the problem that existing voiceprint recognition methods have poor robustness to noisy far-field speech and cannot meet the voiceprint recognition needs of users who have not registered in advance.

[0004] The present invention solves the above problems through the following technical solution:

[0005] A registration-free voiceprint recognition method includes:

[0006] Step S1: Obtain the speech data to be tested and extract the voiceprint features of the speech data to be tested;

[0007] Step S2: Compare the voiceprint features of the voice data to be detected with the voiceprint features of registered users in the database registry. If no matching registered user is found, identify it as a new user and assign a new user ID to the new user. Store the user's voiceprint features and the new user ID in the registry and return the new user ID. If a matching registered user is found, return the user ID of the registered user.

[0008] This invention can perform voiceprint recognition even for users who have not registered in advance, eliminating the cumbersome registration process in existing voiceprint recognition technologies. It also solves the problem that existing voiceprint recognition technologies require the registered voice data of the user to be identified and cannot recognize users who have not recorded their registered voice data, thus enabling users to interact with voice recognition devices more naturally and conveniently.

[0009] It also includes updating the registry using clustering methods and mapping the original user IDs in the registry to the clustering results. The methods for mapping the original user IDs to the clustering results include, but are not limited to, similarity-based mapping methods and statistical result-based mapping methods.

[0010] Clustering methods are used to cluster all collected voice data, and the clustering results are mapped to the original user IDs, with the mapping results saved in the updated registry. By calculating the similarity between the clustered feature centers and the features of the original user IDs, the most similar feature center with a similarity greater than a threshold is selected as the new feature center for each original user ID. Multiple user IDs can correspond to one clustered feature center. The clustering process can learn and correct the original recognition results, and merge data where one person is identified as multiple people. The merged recognition result only reflects the first user ID, and the result is recorded in the database. The recognition efficiency is improved by using clustering.

[0011] The methods for extracting voiceprint features from the speech data to be tested include, but are not limited to, voiceprint feature extraction methods based on Ivctor, voiceprint feature extraction methods based on xvector, and voiceprint feature extraction methods based on ResNet deep learning networks.

[0012] The voiceprint feature extraction model employs a ResNet deep network model based on an attention mechanism. This model uses a ResNet50 network as its base and a deep residual network consisting of four residual network blocks based on the CBAM attention mechanism as its main extraction module. Each residual network block based on the CBAM attention mechanism includes a channel attention module and a spatial attention module, which perform channel and spatial attention respectively. The front end uses 320-dimensional FFT to extract spectral features as input to the deep residual network, and the back end uses the softmax algorithm for speaker-based classification training. The softmax network layer at the back end is removed from the trained model, and the output of the deep residual network is used as the final extracted voiceprint features.

[0013] The methods for comparing the voiceprint features of the speech data to be detected with the voiceprint features of registered users in the database registry include LDA-based feature comparison methods, PLDA-based feature comparison methods, cosine similarity-based feature comparison methods, and attention-based feature similarity comparison methods.

[0014] A registration-free voiceprint recognition system, comprising:

[0015] The main control module communicates with the client and is used to receive client requests and voice data and identify the device ID and account ID in the request; to activate the voice collection module according to the request and forward the client's voice data; to search the database according to the request of the voiceprint recognition module, find the registry corresponding to the specified device ID and user ID and return it; and to return the recognition result of the voiceprint recognition module to the client.

[0016] The voice collection module is used to receive voice data forwarded by the main control module and preprocess the voice data;

[0017] The voiceprint recognition module is used to extract voiceprint features from the preprocessed voice data of the voice collection module; to request the registry of the device ID and account ID corresponding to the voice data from the main control module and to receive the registry returned by the main control module; it is also used to compare the extracted voiceprint features with the voiceprint features in the registry to obtain the similarity, select the user ID with the highest similarity as the identified user ID, if the similarity is greater than a preset threshold, the identified user ID is used as the identification result and fed back to the main control module; if the similarity is less than or equal to the preset threshold, the voiceprint feature is marked as a new user, a new user ID is assigned to the voiceprint feature, and the voiceprint feature is registered in the registry.

[0018] The client's requests and voice data are collected by the voiceprint service module deployed on the client through the recording device after the client is activated, and the voice data and requests are sent to the main control module.

[0019] It also includes a voiceprint feature clustering module, which is used to cluster voiceprint features in the database registry according to the instructions of the main control module, map the clustering results to the registry before clustering, and update the clustering results to the new registry and store them in the database.

[0020] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0021] (1) This invention can also perform voiceprint recognition for users who have not registered in advance, eliminating the cumbersome registration process in the existing voiceprint recognition technology, and solving the problem that the existing voiceprint recognition technology requires the registered voice data of the user to be identified and cannot identify users who have not recorded the registered voice data.

[0022] (2) The present invention expands the existing technology, which can only provide voiceprint-related services to registered users, to provide voiceprint-related services to each member under the account. In particular, in voice usage scenarios with low user registration rates, the voiceprint-related services are significantly improved. In the existing scenario of using personalized voice recommendation, the activation rate of personalized voice recommendation services is increased by more than 150%.

[0023] (3) The present invention uses a clustering method to learn and correct the original recognition results, and maps the speaker IDs before and after clustering to establish the connection between the data before and after clustering. It also merges the data of one person being identified as multiple people. The merged recognition result only feeds back the first user ID and records the result in the database, thereby improving recognition efficiency. Attached Figure Description

[0024] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to embodiments, but the implementation of the present invention is not limited thereto.

[0026] Example 1:

[0027] A registration-free voiceprint recognition method includes:

[0028] Step S1: Obtain the speech data to be tested and extract the voiceprint features of the speech data to be tested;

[0029] Step S2: Compare the voiceprint features of the voice data to be detected with the voiceprint features of registered users in the database registry. If no matching registered user is found, identify it as a new user and assign a new user ID to the new user. Store the user's voiceprint features and the new user ID in the registry and return the new user ID. If a matching registered user is found, return the user ID of the registered user.

[0030] This invention can perform voiceprint recognition even for users who have not registered in advance, eliminating the cumbersome registration process in existing voiceprint recognition technologies. It also solves the problem that existing voiceprint recognition technologies require the registered voice data of the user to be identified and cannot recognize users who have not recorded their registered voice data, thus enabling users to interact with voice recognition devices more naturally and conveniently.

[0031] It also includes updating the registry using clustering methods and mapping the original user IDs in the registry to the clustering results. The methods for mapping the original user IDs to the clustering results include, but are not limited to, similarity-based mapping methods and statistical result-based mapping methods.

[0032] Clustering methods are used to cluster all collected voice data, and the clustering results are mapped to the original user IDs. The mapping results are then saved and updated in the registry. By calculating the similarity between the clustered feature centers and the features of the original user IDs, the most similar feature center with a similarity greater than a threshold is selected as the new feature center for each original user ID. Multiple user IDs can correspond to one clustered feature center. The clustering process can learn and correct the original recognition results, and merge data where one person is identified as multiple people. The merged recognition result only reflects the first user ID, and the result is recorded in the database.

[0033] The experiment verified the following:

[0034] Under a single device and single account, with 10 or fewer voice users, the effective recognition rate for near-field speech (i.e., during testing, the original user ID with the most recognized voices in the results for each registered ID is the effective recognition result for that registration result, and each original user ID corresponds to only one effective recognition result, which is determined by the registered ID with the most recognized voices for that original user ID) increased from 71.1% without using clustering algorithms to 98.6% using clustering and mapping algorithms; the effective recognition rate for far-field speech with environmental noise and reverberation increased from 46.2% to 82.6%.

[0035] The methods for extracting voiceprint features from the speech data to be tested include, but are not limited to, voiceprint feature extraction methods based on Ivctor, voiceprint feature extraction methods based on xvector, and voiceprint feature extraction methods based on ResNet deep learning networks.

[0036] The voiceprint feature extraction model employs a ResNet deep network model based on an attention mechanism. This model uses a ResNet50 network as its base and a deep residual network consisting of four residual network blocks based on the CBAM attention mechanism as its main extraction module. Each residual network block based on the CBAM attention mechanism includes a channel attention module and a spatial attention module, which perform channel and spatial attention respectively. The front end uses 320-dimensional FFT to extract spectral features as input to the deep residual network, and the back end uses the softmax algorithm for speaker-based classification training. The softmax network layer at the back end is removed from the trained model, and the output of the deep residual network is used as the final extracted voiceprint features.

[0037] The methods for comparing the voiceprint features of the speech data to be detected with the voiceprint features of registered users in the database registry include LDA-based feature comparison methods, PLDA-based feature comparison methods, cosine similarity-based feature comparison methods, and attention-based feature similarity comparison methods.

[0038] Example 2:

[0039] Combined with appendix Figure 1 As shown, a registration-free voiceprint recognition system includes:

[0040] The main control module communicates with the client. The voiceprint service module deployed on the client is activated by the client and collects voice data through the recording device. The client then sends the voice data and requests to the main control module. The main control module identifies the device ID and account ID in the request. It is used to activate the voice collection module according to the request and forward the client's voice data. It is also used to search the database according to the request of the voiceprint recognition module, find the registry corresponding to the specified device ID and user ID and return it. Finally, it is used to return the recognition result of the voiceprint recognition module to the client.

[0041] The voice collection module is used to receive voice data forwarded by the main control module and preprocess the voice data;

[0042] The voiceprint recognition module is used to extract voiceprint features from the preprocessed voice data of the voice collection module; to request the registry of the device ID and account ID corresponding to the voice data from the main control module and to receive the registry returned by the main control module; it is also used to compare the extracted voiceprint features with the voiceprint features in the registry to obtain the similarity, select the user ID with the highest similarity as the identified user ID, if the similarity is greater than a preset threshold, the identified user ID is used as the identification result and fed back to the main control module; if the similarity is less than or equal to the preset threshold, the voiceprint feature is marked as a new user, a new user ID is assigned to the voiceprint feature, and the voiceprint feature is registered in the registry.

[0043] The voiceprint recognition process is as follows:

[0044] First, the main control module loads and activates the system, establishes a communication connection with the client, and receives the client's unregistered voiceprint recognition service application and the audio data to be processed. Then, the voice collection module collects the audio data received by the main control module, preprocesses the audio data, and hands it over to the voiceprint recognition module for feature extraction and comparison, as well as speaker identification. If the voice recognition identifies a stranger, the voice feature is registered, a new user ID is assigned to it, and it is recorded in the registry of the database. The recognition result is then fed back to the main control module, which sends the feedback result to the client. This invention can efficiently complete voiceprint recognition tasks without registration. The registration-free voiceprint recognition method eliminates the cumbersome registration process in voiceprint recognition, allowing users to interact with devices more naturally and conveniently, providing crucial user identification capabilities for personalized recommendation systems, personalized service systems, and more.

[0045] It also includes a voiceprint feature clustering module, which is used to cluster voiceprint features in the database registry according to the instructions of the main control module, map the clustering results to the registry before clustering, and update the clustering results to the new registry and store them in the database.

[0046] The voiceprint clustering process is as follows: First, the main control module activates the voiceprint clustering task at regular intervals. Then, the voiceprint feature clustering module performs voiceprint feature clustering on the existing feature data in the database, saves the clustering results, updates the clustering results to user registration data, and establishes a mapping relationship with the original user ID.

[0047] Although the present invention has been described herein with reference to illustrative embodiments, the above embodiments are merely preferred embodiments of the present invention, and the implementation of the present invention is not limited to the above embodiments. It should be understood that those skilled in the art can devise many other modifications and implementations, which will fall within the scope and spirit of the principles disclosed in this application.

Claims

1. A registration-free voiceprint recognition method, characterized in that, This is implemented based on a registration-free voiceprint recognition system, which includes: The main control module communicates with the client and is used to receive client requests and voice data and identify the device ID and account ID in the request; to activate the voice collection module according to the request and forward the client's voice data; to search the database according to the request of the voiceprint recognition module, find the registry corresponding to the specified device ID and user ID and return it; and to return the recognition result of the voiceprint recognition module to the client. The voice collection module is used to receive voice data forwarded by the main control module and preprocess the voice data; The voiceprint recognition module is used to extract voiceprint features from the preprocessed voice data of the voice collection module; to request the registry of the device ID and account ID corresponding to the voice data from the main control module and to receive the registry returned by the main control module; it is also used to compare the extracted voiceprint features with the voiceprint features in the registry to obtain the similarity, select the user ID with the highest similarity as the identified user ID, if the similarity is greater than a preset threshold, the identified user ID is used as the identification result and fed back to the main control module; if the similarity is less than or equal to the preset threshold, the voiceprint feature is marked as a new user, a new user ID is assigned to the voiceprint feature, and the voiceprint feature is registered in the registry. It also includes a voiceprint feature clustering module, which is used to cluster the voiceprint features in the database registry according to the instructions of the main control module, establish a mapping between the clustering results and the original user ID in the registry before clustering, and update the clustering results to the new registry and store them in the database. The methods include: Step S1: Obtain the speech data to be tested and extract the voiceprint features of the speech data to be tested; Step S2: Compare the voiceprint features of the voice data to be detected with the voiceprint features of registered users in the database registry. If no matching registered user is found, identify it as a new user and assign a new user ID to the new user. Store the user's voiceprint features and the new user ID in the registry and return the new user ID. If a matching registered user is found, return the user ID of the registered user. The method also includes updating the registry using a clustering approach and mapping the original user IDs in the registry to the clustering results.

2. The registration-free voiceprint recognition method according to claim 1, characterized in that, Methods for mapping original user IDs to clustering results include similarity-based mapping methods and statistical result-based mapping methods.

3. The registration-free voiceprint recognition method according to claim 1, characterized in that, The methods for extracting voiceprint features from the speech data to be tested include an ivector-based voiceprint feature extraction method, an xvector-based voiceprint feature extraction method, and a ResNet deep learning network-based voiceprint feature extraction method.

4. The registration-free voiceprint recognition method according to claim 1, characterized in that, The voiceprint feature extraction model employs a ResNet deep network model based on an attention mechanism. This model uses a ResNet50 network as its base and a deep residual network consisting of four residual network blocks based on the CBAM attention mechanism as its main extraction module. Each residual network block based on the CBAM attention mechanism includes a channel attention module and a spatial attention module, which perform channel and spatial attention respectively. The front end uses 320-dimensional FFT to extract spectral features as input to the deep residual network, and the back end uses the softmax algorithm for speaker-based classification training. The softmax network layer at the back end is removed from the trained model, and the output of the deep residual network is used as the final extracted voiceprint features.

5. The registration-free voiceprint recognition method according to claim 1, characterized in that, The methods for comparing the voiceprint features of the speech data to be detected with the voiceprint features of registered users in the database registry include LDA-based feature comparison methods, PLDA-based feature comparison methods, cosine similarity-based feature comparison methods, and attention-based feature similarity comparison methods.

Images