Speaker recognition model training, speaker recognition method and device

Abstract

The present application discloses a speaker recognition model training, a speaker recognition method and a device, and relates to the field of smart home technology. The speaker recognition model training method includes: obtaining a sample speech set; the sample speech set includes: N sample speech corresponding to multiple users; extracting acoustic features of each sample speech to obtain N initial sample speech acoustic features corresponding to multiple users; reducing the dimension of the N initial sample speech acoustic features through a target feature extraction model to obtain M target sample speech acoustic features corresponding to multiple users; M is an integer less than N and greater than or equal to 2; clustering the M target sample speech acoustic features to obtain clustering results; the clustering results are used to characterize the users to which each target sample speech acoustic feature belongs; according to the clustering results, and the sample speech set, a preset model is trained to obtain a speaker recognition model. The present application improves the efficiency of speaker recognition model training.

Description

Technical Field

[0001] The present application relates to the field of smart home, and more specifically, to a speaker recognition model training, a speaker recognition method and a device. Background Art

[0002] Taking smart washing machines, smart water heaters and other smart homes as examples, smart homes can perform speaker recognition based on the speaker's voice and perform related operations based on the speaker recognition results. At present, the common speaker recognition method is mainly based on deep learning. Before using this speaker recognition method for speaker recognition, it is necessary to use a large amount of sample voice data annotated with user labels to train the deep learning model to obtain a model that can be used for speaker recognition.

[0003] In the existing speaker recognition model training method, it is necessary to manually label the above-mentioned sample speech data to obtain the sample speech data labeled with user labels. Therefore, the existing speaker recognition model training method has the problem of low efficiency. Summary of the invention

[0004] The present application provides a speaker recognition model training, a speaker recognition method and a device to improve the efficiency of speaker recognition model training.

[0005] In a first aspect, the present application provides a speaker recognition model training method, the method comprising:

[0006] Acquire a sample voice set; the sample voice set includes: N sample voices corresponding to multiple users; N is greater than or equal to 2;

[0007] Extracting acoustic features from each of the sample speech to obtain acoustic features of N initial sample speech corresponding to the multiple users;

[0008] By using a target feature extraction model, the N initial sample speech acoustic features are reduced in dimension to obtain M target sample speech acoustic features corresponding to the multiple users; wherein M is an integer less than N and greater than or equal to 2;

[0009] Clustering the M target sample speech acoustic features to obtain a clustering result; the clustering result is used to characterize the user to which each of the target sample speech acoustic features belongs;

[0010] According to the clustering result and the sample speech set, a preset model is trained to obtain a speaker recognition model; the speaker recognition model is used to output an identifier of a user corresponding to the target speech based on the target speech.

[0011] Optionally, the obtaining of a speaker recognition model by training a preset model according to the clustering result and the sample speech set includes:

[0012] Determining a user label for each sample speech in the sample speech set according to the user to which the acoustic features of each target sample speech belong;

[0013] The sample speech set and the user label of each sample speech are used to train a preset model to obtain a speaker recognition model.

[0014] Optionally, clustering the M target sample speech acoustic features to obtain a clustering result includes:

[0015] By using a preset clustering algorithm, the M target sample speech acoustic features are clustered into an indefinite number of categories to obtain multiple groups of target sample speech acoustic features; wherein the target sample speech acoustic features in the same group correspond to the same user;

[0016] A user label corresponding to each group of the target sample speech acoustic features is received to obtain the clustering result.

[0017] Optionally, before the target feature extraction model is used to reduce the dimension of the N initial sample speech acoustic features, the method further includes:

[0018] The preset unsupervised feature extraction model is trained using the N initial sample speech acoustic features to obtain the target feature extraction model.

[0019] Optionally, the loss function of the preset unsupervised feature extraction model is shown in the following formula (1):

[0020]

[0021] Wherein, L represents the loss function of the preset unsupervised feature extraction model, W represents the weight matrix of the preset unsupervised feature extraction model, c is a constant, y represents any of the initial sample speech acoustic features, o() represents the sigmoid function, || 1 represents the L1 norm, W T represents the transpose of the weight matrix, || 2 represents the L2 norm.

[0022] In a second aspect, the present application provides a speaker recognition method, the method comprising:

[0023] Collect the user's target voice;

[0024] Using a speaker recognition model, performing speaker recognition on the target speech, and outputting an identifier of a user corresponding to the target speech; wherein the speaker recognition model is trained using the method described in any one of the first aspects;

[0025] According to the user identifier, a target operation corresponding to the user is executed.

[0026] In a third aspect, the present application provides a speaker recognition model training device, the device comprising:

[0027] An acquisition module, used to acquire a sample voice set; the sample voice set includes: N sample voices corresponding to multiple users; N is greater than or equal to 2;

[0028] The first processing module is used to extract acoustic features of each of the sample speech to obtain N initial sample speech acoustic features corresponding to the multiple users; through the target feature extraction model, the N initial sample speech acoustic features are reduced in dimension to obtain M target sample speech acoustic features corresponding to the multiple users; M is an integer less than N and greater than or equal to 2;

[0029] A clustering module, used for clustering the M target sample speech acoustic features to obtain a clustering result; the clustering result is used to characterize the user to which each of the target sample speech acoustic features belongs;

[0030] The second processing module is used to train a preset model according to the clustering result and the sample speech set to obtain a speaker recognition model; the speaker recognition model is used to output an identifier of a user corresponding to the target speech based on the target speech.

[0031] In a fourth aspect, the present application provides a speaker recognition device, the device comprising:

[0032] A collection module, used to collect the target voice of the user;

[0033] a recognition module, configured to use a speaker recognition model to perform speaker recognition on the target speech and output an identifier of a user corresponding to the target speech; wherein the speaker recognition model is trained using the method described in any one of the first aspects;

[0034] The execution module is used to execute the target operation corresponding to the user according to the identifier of the user.

[0035] In a fifth aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium includes a stored program, wherein when the program is run, the method described in any one of the first aspect and the second aspect is executed.

[0036] In a sixth aspect, the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute any one of the methods in the first aspect and the second aspect through the computer program.

[0037] The speaker recognition model training, speaker recognition method and device provided by the present application can obtain the initial sample speech acoustic features corresponding to each sample speech by extracting the acoustic features of the sample speech. By reducing the dimension of N non-labeled initial sample speech acoustic features, the data volume of the sample speech acoustic features to be clustered is reduced, the clustering efficiency is improved, and the efficiency of speaker recognition model training is improved. By clustering the target sample speech acoustic features after dimensionality reduction, the user to which each target sample speech acoustic feature belongs can be determined. According to the above clustering results and sample speech, the preset model can be trained to obtain a speaker recognition model for speaker recognition. Through the above method, there is no need to manually label a large amount of sample speech data, which improves the efficiency of obtaining data for training the preset model, thereby further improving the efficiency of speaker recognition model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0039] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0040] Figure 1 A schematic diagram of an application scenario of a speaker recognition method;

[0041] Figure 2 A flowchart of a speaker recognition model training method provided in this application;

[0042] Figure 3 A flowchart of a speaker recognition method provided in this application;

[0043] Figure 4 A schematic diagram of the structure of a speaker recognition model training device provided in this application;

[0044] Figure 5 A schematic diagram of the structure of a speaker recognition device provided in this application;

[0045] Figure 6A schematic diagram of the structure of an electronic device provided in this application. DETAILED DESCRIPTION

[0046] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0047] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0048] Figure 1 Schematic diagram of an application scenario of a speaker recognition method. Figure 1 As shown, speaker recognition can be widely used in smart home, smart home, smart home equipment ecology, smart residential (IntelligenceHouse) ecology and other whole-house intelligent digital control application scenarios. Optionally, the above speaker recognition method can be applied to Figure 1 In the hardware environment composed of the terminal device 12 and the server 14 shown in FIG. Figure 1 As shown, the server 14 is connected to the terminal device 12 via a network, and can be used to provide services (such as application services, etc.) for the terminal or a client installed on the terminal. A database can be set on the server or independently of the server to provide data storage services for the server 14. Cloud computing and / or edge computing services can be configured on the server or independently of the server to provide data computing services for the server 14.

[0049] The above network may include but is not limited to at least one of the following: wired network, wireless network. The above wired network may include but is not limited to at least one of the following: wide area network, metropolitan area network, local area network, and the above wireless network may include but is not limited to at least one of the following: WIFI (Wireless Fidelity), Bluetooth. The terminal device 12 may be but is not limited to a PC, a mobile phone, a tablet computer, a smart air conditioner, a smart range hood, a smart refrigerator, a smart oven, a smart stove, a smart washing machine, a smart water heater, a smart washing equipment, a smart dishwasher, a smart projection equipment, a smart TV, a smart clothes drying rack, a smart curtain, a smart audio and video, a smart socket, a smart speaker, a smart speaker, a smart fresh air equipment, a smart kitchen and bathroom equipment, a smart bathroom equipment, a smart sweeping robot, a smart window cleaning robot, a smart mopping robot, a smart air purification equipment, a smart steamer, a smart microwave oven, a smart kitchen treasure, a smart purifier, a smart water dispenser, a smart door lock, etc.

[0050] Taking smart home as an example, the smart home collects the speaker's voice and sends it to the server. The server can perform speaker recognition and control the smart home to perform related operations based on the speaker recognition results. At present, the common speaker recognition method is mainly based on deep learning. Before using this speaker recognition method for speaker recognition, it is necessary to use a large amount of sample speech data marked with user labels (that is, speaker identity information) to train the deep learning model with a supervised classifier to obtain a model that can be used for speaker recognition.

[0051] In the existing speaker recognition model training method, it is necessary to manually label the above-mentioned sample speech data to obtain sample speech data labeled with user labels. However, the process of labeling a large amount of sample speech data with user labels is cumbersome and requires a lot of manual time. Therefore, the efficiency of the existing speaker recognition model training method is low.

[0052] Considering the above-mentioned low efficiency problem of the existing speaker recognition model, this application proposes a method for training a preset model to obtain a speaker recognition model without manually annotating a large amount of sample speech data. Through the above method, automated model training is achieved to obtain a speaker recognition model, avoiding manual annotation of a large amount of sample speech data, thereby improving the efficiency of speaker recognition model training.

[0053] The speaker recognition model training method provided by the present application is described in detail below in conjunction with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. It should be understood that the executor of the speaker recognition model training method can be any electronic device such as a terminal or server with processing functions.

[0054] Figure 2 A flowchart of a speaker recognition model training method provided in this application. Figure 2 As shown, the method comprises the following steps:

[0055] S101: Obtain a sample speech set.

[0056] The sample speech set may include: N sample speech corresponding to multiple users. Where N is greater than or equal to 2. It should be understood that the present application does not limit the time length of the sample speech. In addition, the time lengths of different sample speech can be the same or different. For example, the sample speech set can be obtained by X u ={x (1) u , x (2) u , …, x (mu) u}, where u represents that data is not labeled, and mu represents the data sequence number (that is, there are mu sample voices in the sample voice set).

[0057] It should be understood that the present application does not limit how the electronic device obtains the above-mentioned sample voice set. Optionally, the electronic device may receive the sample voice set input by the user through an application programming interface (API) or a graphical user interface (GUI). Alternatively, the above-mentioned sample voice set may also be pre-stored in the electronic device. The electronic device may read the sample voice set from the electronic device according to the storage path of the sample voice set. Alternatively, the electronic device may also obtain the sample voice set from other devices that store the sample voice set.

[0058] Alternatively, the electronic device may also be provided with a voice collection device, through which the electronic device may collect N sample voices corresponding to the above-mentioned multiple users. In this implementation, the electronic device may be the same type of electronic device as the electronic device that performs the following speaker recognition method, for example, the electronic device and the electronic device that performs the following speaker recognition method may both be smart home devices.

[0059] S102: Extract acoustic features from each sample speech to obtain N initial sample speech acoustic features corresponding to multiple users.

[0060] Optionally, the above-mentioned initial sample speech acoustic features may include, for example, at least one of the following: Mel-Frequency Cepstral Coefficients (MFCC) features, filter group fbank features, etc. Exemplarily, taking the initial sample speech acoustic features including MFCC features and fbank features as an example, the electronic device may, for example, obtain the MFCC features of the sample speech by an acoustic feature extraction method corresponding to the MFCC features, and obtain the fbank features of the sample speech by an acoustic feature extraction method corresponding to the fbank features, and then feature splice the MFCC features and fbank features of the sample speech to obtain the initial sample speech acoustic features corresponding to the sample speech.

[0061] It should be understood that, for any sample speech, the electronic device can obtain the initial sample speech acoustic features corresponding to the sample speech by any existing acoustic feature extraction method, which will not be elaborated here.

[0062] For example, taking N equal to mu as an example, the N initial sample speech acoustic features corresponding to the multiple users can be obtained by Y u ={y (1) u ,y (2) u , …, y (mu) u}. Among them, u means that the data is not labeled.

[0063] S103. Using a target feature extraction model, reduce the dimension of the N initial sample speech acoustic features to obtain M target sample speech acoustic features corresponding to multiple users.

[0064] Wherein, the above M is an integer less than N and greater than or equal to 2. The electronic device can input the above N initial sample speech acoustic features into the target feature extraction model to obtain M target sample speech acoustic features corresponding to multiple users. Optionally, the target feature extraction model can be pre-stored in the electronic device by the user.

[0065] For example, taking the above M equal to nu as an example, the M target sample speech acoustic features corresponding to the above multiple users can be obtained by A u ={a (1) u , a (2) u , …, a (nu) u}. Among them, u means that the data is not labeled.

[0066] S104: Cluster the M target sample speech acoustic features to obtain clustering results.

[0067] The above clustering results are used to characterize the user to which each target sample speech acoustic feature belongs.

[0068] As a possible implementation, the electronic device may cluster the M target sample speech acoustic features into an indefinite number of categories using a preset clustering algorithm to obtain multiple groups of target sample speech acoustic features, wherein the target sample speech acoustic features in the same group correspond to the same user.

[0069] Exemplarily, the preset clustering algorithm may be, for example, a Gaussian Mixture Model (GMM) clustering algorithm, or a k-means clustering algorithm, etc. The GMM clustering algorithm may fit the M target sample speech acoustic features into the number of categories with the maximum total likelihood, so that the probability that the target sample speech acoustic features of the same user belong to the same category is higher, thereby realizing clustering of the target sample speech acoustic features.

[0070] After obtaining multiple groups of target sample speech acoustic features, the electronic device can receive a user tag corresponding to each group of target sample speech acoustic features to obtain the above clustering result. Optionally, the electronic device can receive user tags for each group of target sample speech acoustic features input by the user. According to the user tags corresponding to each group, the electronic device can determine the user to which the group of target sample speech acoustic features belongs, and then obtain the above clustering result.

[0071] In this implementation, assuming that 5 groups of target sample speech acoustic features are obtained through clustering, only 5 user labels need to be manually input. Compared with the prior art that requires manual labeling of all sample speech, this application reduces the workload of manual labeling and improves the efficiency of electronic devices in obtaining sample sets for training speaker recognition models, thereby improving the training efficiency of speaker recognition models.

[0072] S105. According to the above clustering results and the sample speech set, a preset model is trained to obtain a speaker recognition model.

[0073] The speaker recognition model is used to output the identifier of the user corresponding to the target speech based on the target speech.

[0074] Optionally, the present application does not limit the above preset model. Exemplarily, the preset model can be, for example, a generative classification model, such as a probabilistic linear discriminant analysis (PLDA) model. In some embodiments, based on the above clustering results and sample speech sets, a voiceprint feature extraction model can also be trained to obtain a trained voiceprint feature extraction model for voice signal voiceprint feature extraction.

[0075] As a possible implementation method, after the electronic device obtains the above clustering results, in the above clustering results, each target sample speech acoustic feature belonging to the same cluster cluster corresponds to the same user, and each target sample speech acoustic feature not belonging to the same cluster cluster corresponds to different users. For example, if the clustering result is 3 clusters, it can be determined that there are also 3 users; each cluster corresponds to user A, user B and user C respectively. Then, there is no need to annotate each sample speech in the sample speech set to achieve the purpose of identifying the voiceprints of different users, which reduces the success of manual annotation and improves the efficiency of speaker recognition.

[0076] In this embodiment, the user to which each target sample speech acoustic feature belongs can be determined based on the clustering result, as well as the user label corresponding to each sample speech in the sample speech set. In some embodiments, the user label can be determined based on the recognized speech acoustic features, for example, the user label of each sample speech is set to adult male, adult female, child, elderly, etc. based on the speech acoustic features; or it can also be marked according to the set label, for example, the user label of each sample speech is set to user A, user B, user C, etc., to achieve the purpose of distinguishing the voice commands of each user. Then, the electronic device can use the aforementioned sample speech set and the user label of each sample speech to train the above preset model to obtain the speaker recognition model.

[0077] Optionally, the electronic device can determine the user to which each initial sample speech acoustic feature belongs based on the user to which each target sample speech acoustic feature belongs. Optionally, the electronic device can use the above-mentioned target feature extraction model to combine multiple initial sample speech acoustic features to obtain a target sample speech acoustic feature, and then implement dimensionality reduction of the initial sample speech acoustic feature. The electronic device can also record multiple initial sample speech acoustic features before dimensionality reduction corresponding to each target sample speech acoustic feature. In this implementation mode, the electronic device can use the user to which a target sample speech acoustic feature belongs as the user to which multiple initial sample speech acoustic features before dimensionality reduction corresponding to the target sample speech acoustic feature belong.

[0078] The initial sample speech acoustic features correspond to the sample speech one-to-one, so the electronic device can determine the user to whom the initial sample speech acoustic features belong as the user to which the sample speech belongs, and further the electronic device can determine the user label of each sample speech.

[0079] In this embodiment, by extracting acoustic features from sample speech, the initial sample speech acoustic features corresponding to each sample speech can be obtained. By reducing the dimension of N non-labeled initial sample speech acoustic features, the amount of data of the sample speech acoustic features to be clustered is reduced, the clustering efficiency is improved, and the efficiency of speaker recognition model training is improved. By clustering the target sample speech acoustic features after dimensionality reduction, the user to which each target sample speech acoustic feature belongs can be determined. According to the above clustering results and sample speech, the preset model can be trained to obtain a speaker recognition model for speaker recognition. Through the above method, there is no need to manually label a large amount of sample speech data, which improves the efficiency of obtaining data for training the preset model, and further improves the efficiency of speaker recognition model training.

[0080] The following is a detailed description of how electronic devices obtain target feature extraction models:

[0081] As a possible implementation method, before the electronic device reduces the dimension of the N initial sample speech acoustic features through the target feature extraction model, it can also use the above N initial sample speech acoustic features to train a preset unsupervised feature extraction model to obtain the target feature extraction model.

[0082] It should be understood that the present application does not limit the type of the above-mentioned preset unsupervised feature extraction model. Optionally, the preset unsupervised feature extraction model can be, for example, any existing model that can achieve data dimensionality reduction based on sample speech that has not been labeled with user labels. For example, an autoencoding dimensionality reduction model based on BP neural network (Back Propagation Neural Network, BPNN), or an autoencoding dimensionality reduction model based on convolutional neural networks (Convolutional Neural Networks, CNN), etc.

[0083] In this implementation, optionally, the loss function of the above-mentioned preset unsupervised feature extraction model is, for example, as shown in the following formula (1):

[0084]

[0085] Among them, L represents the loss function of the preset unsupervised feature extraction model, W represents the weight matrix of the preset unsupervised feature extraction model, c is a constant, y represents the acoustic feature of any initial sample speech, o() represents the sigmoid function, ||1 represents the L1 norm, W T represents the transpose of the weight matrix, || 2 Represents the L2 norm. Among them, the L1 norm refers to the sum of the absolute values ​​of each element in the vector. The L2 norm refers to the result of taking the square root of the sum of the squares of all the elements of the vector. Through the above loss function, there is no need to calculate the value of the loss function in the process of training the preset unsupervised feature extraction model according to the user label of each sample speech, which realizes the unsupervised training of the preset unsupervised feature extraction model, avoids manual labeling of the sample speech, and further improves the efficiency of training the aforementioned preset model to obtain the speaker recognition model.

[0086] The input of the exemplary preset unsupervised feature extraction model can be the initial sample speech acoustic features, and the output is the target sample speech acoustic features. Optionally, the training process of the preset unsupervised feature extraction model can be iteratively optimized using a gradient descent algorithm, and after the iterative optimization is completed, a trained target feature extraction model is obtained. Exemplarily, the target feature extraction model can be used to reduce the dimension of N initial sample speech acoustic features to obtain M target sample speech acoustic features.

[0087] As another possible implementation manner, the above-mentioned target feature extraction model may also be pre-stored in the electronic device by the user.

[0088] In this embodiment, the preset unsupervised feature extraction model can be unsupervisedly trained through the acoustic features of N initial sample voices, and then the target feature extraction model can be obtained. Through unsupervised training, it is not necessary to train the preset unsupervised feature extraction model according to the user label of each sample voice, which avoids manual labeling of the sample voice and further improves the efficiency of obtaining the speaker recognition model by training the aforementioned preset model.

[0089] After obtaining the above speaker recognition model, the speaker recognition model can be used to perform speaker recognition. It should be understood that the execution subject of the following speaker recognition method can be any electronic device with processing capabilities. Optionally, the electronic device and the electronic device for executing the above method for executing the speaker recognition model can be the same electronic device or different electronic devices, and this application does not limit this. Exemplarily, the electronic device executing the speaker recognition method can be, for example, the above Figure 1 The terminal device 12 or the server 14 shown in FIG.

[0090] The speaker recognition method provided by the present application is described in detail below in conjunction with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0091] Figure 3 A flow chart of a speaker recognition method provided in this application. Figure 3 As shown, the method comprises the following steps:

[0092] S201: Collect the target voice of the user.

[0093] Optionally, the electronic device may collect the target voice of the user through a voice collection device, wherein the voice collection device may be integrated in the electronic device, or the voice collection device may be a voice collection device independent of the electronic device.

[0094] It should be understood that the present application does not limit the time length of the target speech collected by the electronic device.

[0095] S202: Use a speaker recognition model to perform speaker recognition on the target speech, and output an identifier of a user corresponding to the target speech.

[0096] The speaker recognition model is trained by using the method described in any of the above embodiments.

[0097] In some embodiments, before using the above speaker recognition model to recognize the target voice, the electronic device may also use a trained voiceprint feature extraction model to extract voiceprint features from the target voice to obtain voiceprint features corresponding to the target voice. Then, the electronic device may input the voiceprint features corresponding to the target voice into the above speaker recognition model to obtain the identifier of the user corresponding to the target voice. The above trained voiceprint feature extraction model may be trained based on the above clustering results and sample voice set.

[0098] S203: Execute the target operation corresponding to the user according to the user's identification.

[0099] It should be understood that the present application does not limit the target operation corresponding to the above-mentioned user. For example, the electronic device can implement voice unlocking or voice wake-up of a specific person according to the user's identification.

[0100] In this embodiment, after collecting the target voice of the user, the electronic device can use the speaker recognition model obtained by training the speaker recognition model described in any of the above embodiments to perform speaker recognition on the target voice, thereby obtaining the identification of the user corresponding to the target voice. The present application trains the speaker recognition model by a method that does not require a large amount of manual annotation of sample voices, thereby improving the efficiency of obtaining the speaker recognition model, thereby also improving the efficiency of obtaining the user identification by performing speaker recognition based on the speaker recognition model, thereby improving the intelligence of speaker recognition and expanding the scope of application.

[0101] Figure 4This is a schematic diagram of the structure of a speaker recognition model training device provided in this application. Figure 4 As shown, the device 30 includes: an acquisition module 31, a first processing module 32, a clustering module 33, and a second processing module 34.

[0102] The acquisition module 31 is used to acquire a sample speech set, wherein the sample speech set includes: N sample speech corresponding to multiple users; wherein N is greater than or equal to 2.

[0103] The first processing module 32 is used to extract acoustic features from each of the sample speech to obtain N initial sample speech acoustic features corresponding to the multiple users; and to reduce the dimension of the N initial sample speech acoustic features through a target feature extraction model to obtain M target sample speech acoustic features corresponding to the multiple users. Wherein, M is an integer less than N and greater than or equal to 2.

[0104] The clustering module 33 is used to cluster the M target sample speech acoustic features to obtain a clustering result, wherein the clustering result is used to characterize the user to which each target sample speech acoustic feature belongs.

[0105] The second processing module 34 is used to train a preset model according to the clustering result and the sample speech set to obtain a speaker recognition model, wherein the speaker recognition model is used to output the identifier of the user corresponding to the target speech based on the target speech.

[0106] Optionally, the second processing module 34 is specifically used to determine the user label of each sample speech in the sample speech set according to the user to which the acoustic features of each target sample speech belong; use the sample speech set and the user label of each sample speech to train a preset model to obtain a speaker recognition model.

[0107] Optionally, the clustering module 33 is specifically used to cluster the M target sample speech acoustic features into an uncertain number of categories through a preset clustering algorithm to obtain multiple groups of target sample speech acoustic features. Among them, the target sample speech acoustic features in the same group correspond to the same user. In this implementation, the device may also include a receiving module 35, which is used to receive a user label corresponding to each group of the target sample speech acoustic features to obtain the clustering result.

[0108] Optionally, the second processing module 34 is also used to train a preset unsupervised feature extraction model using the N initial sample speech acoustic features before reducing the dimension of the N initial sample speech acoustic features through the target feature extraction model to obtain the target feature extraction model.

[0109] Optionally, the loss function of the preset unsupervised feature extraction model is shown in the following formula (1):

[0110]

[0111] Wherein, L represents the loss function of the preset unsupervised feature extraction model, W represents the weight matrix of the preset unsupervised feature extraction model, c is a constant, y represents any of the initial sample speech acoustic features, o() represents the sigmoid function, || 1 represents the L1 norm, W T represents the transpose of the weight matrix, || 2 represents the L2 norm.

[0112] The speaker recognition model training device 30 provided in the present application is used to execute the aforementioned speaker recognition model training method embodiment, and its implementation principle and technical effect are similar, which will not be described in detail.

[0113] Figure 5 This is a schematic diagram of the structure of a speaker recognition device provided by this application. Figure 5 As shown, the device 40 includes:

[0114] The collection module 41 is used to collect the target voice of the user.

[0115] The recognition module 42 is used to perform speaker recognition on the target speech using a speaker recognition model and output an identifier of a user corresponding to the target speech. The speaker recognition model is trained using the method described in any of the above speaker recognition model training embodiments.

[0116] The execution module 43 is used to execute the target operation corresponding to the user according to the user's identification.

[0117] The speaker recognition device 40 provided in the present application is used to execute the above-mentioned speaker recognition method embodiment, and its implementation principle and technical effect are similar, which will not be described in detail.

[0118] Figure 6 This is a schematic diagram of the structure of an electronic device provided in this application. Figure 6 As shown, the electronic device 500 may include: at least one processor 501 and a memory 502 .

[0119] The memory 502 is used to store programs. Specifically, the programs may include program codes, and the program codes include computer operation instructions.

[0120] The memory 502 may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory.

[0121] The processor 501 is used to execute the computer-executable instructions stored in the memory 502 to implement the speaker recognition model training and speaker recognition method described in the aforementioned method embodiment. The processor 501 may be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0122] Optionally, the electronic device 500 may further include a communication interface 503. In a specific implementation, if the communication interface 503, the memory 502 and the processor 501 are implemented independently, the communication interface 503, the memory 502 and the processor 501 may be interconnected through a bus and communicate with each other. The bus may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc., but it does not mean that there is only one bus or one type of bus.

[0123] Optionally, in a specific implementation, if the communication interface 503, the memory 502 and the processor 501 are integrated on a chip, the communication interface 503, the memory 502 and the processor 501 can complete communication through an internal interface.

[0124] The present application also provides a computer-readable storage medium, which may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, and other media that can store program codes. Specifically, the computer-readable storage medium stores program instructions, and the program instructions are used for the methods in the above embodiments.

[0125] The present application also provides a program product, which includes an execution instruction, which is stored in a readable storage medium. At least one processor of an electronic device can read the execution instruction from the readable storage medium, and at least one processor executes the execution instruction so that the electronic device implements the speaker recognition model training and speaker recognition method provided by the various embodiments described above.

[0126] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A speaker recognition model training method, characterized in that: The method comprises: Acquire a sample voice set; the sample voice set includes: N sample voices corresponding to multiple users; N is greater than or equal to 2; Extracting acoustic features of each of the sample speech to obtain N initial sample speech acoustic features corresponding to the multiple users; performing dimension reduction on the N initial sample speech acoustic features through a target feature extraction model to obtain M target sample speech acoustic features corresponding to the multiple users; wherein M is an integer less than N and greater than or equal to 2; The M target sample speech acoustic features are clustered into an indefinite number of categories by a preset clustering algorithm to obtain multiple groups of target sample speech acoustic features; wherein the target sample speech acoustic features in the same group correspond to the same user; a user label corresponding to each group of the target sample speech acoustic features is received to obtain a clustering result; the clustering result is used to characterize the user to which each target sample speech acoustic feature belongs; wherein the preset clustering algorithm is a Gaussian mixture model clustering algorithm; According to the clustering result and the sample speech set, a preset model is trained to obtain a speaker recognition model; the speaker recognition model is used to output an identifier of a user corresponding to the target speech based on the target speech; Before the target feature extraction model is used to reduce the dimension of the N initial sample speech acoustic features, the method further includes: Using the N initial sample speech acoustic features, training a preset unsupervised feature extraction model to obtain the target feature extraction model; The loss function of the preset unsupervised feature extraction model is shown in the following formula (1): （1） in, represents the loss function of the preset unsupervised feature extraction model, represents the weight matrix of the preset unsupervised feature extraction model, is a constant, represents any of the initial sample speech acoustic features, represents the sigmoid function, represents the L1 norm, represents the transpose of the weight matrix, represents the L2 norm.

2. The method according to claim 1, characterized in that The method of training a preset model based on the clustering result and the sample speech set to obtain a speaker recognition model includes: Determining a user label for each sample speech in the sample speech set according to the user to which the acoustic features of each target sample speech belong; The sample speech set and the user label of each sample speech are used to train a preset model to obtain a speaker recognition model.

3. A speaker recognition method, characterized in that: The method comprises: Collect the user's target voice; Using a speaker recognition model, performing speaker recognition on the target speech, and outputting an identifier of a user corresponding to the target speech; wherein the speaker recognition model is trained using the method according to claim 1 or 2; According to the user identifier, a target operation corresponding to the user is executed.

4. A speaker recognition model training device, characterized in that: The device comprises: An acquisition module, used to acquire a sample voice set; the sample voice set includes: N sample voices corresponding to multiple users; N is greater than or equal to 2; The first processing module is used to extract acoustic features of each of the sample speech to obtain N initial sample speech acoustic features corresponding to the multiple users; through the target feature extraction model, the N initial sample speech acoustic features are reduced in dimension to obtain M target sample speech acoustic features corresponding to the multiple users; M is an integer less than N and greater than or equal to 2; A clustering module, used to cluster the M target sample speech acoustic features into an indefinite number of categories by a preset clustering algorithm to obtain multiple groups of target sample speech acoustic features; wherein the target sample speech acoustic features in the same group correspond to the same user; receive a user label corresponding to each group of the target sample speech acoustic features to obtain a clustering result; the clustering result is used to characterize the user to which each of the target sample speech acoustic features belongs; wherein the preset clustering algorithm is a Gaussian mixture model clustering algorithm; A second processing module is used to train a preset model according to the clustering result and the sample speech set to obtain a speaker recognition model; the speaker recognition model is used to output an identifier of a user corresponding to the target speech based on the target speech; The second processing module is further used to train a preset unsupervised feature extraction model using the N initial sample speech acoustic features before reducing the dimension of the N initial sample speech acoustic features through the target feature extraction model to obtain the target feature extraction model; The loss function of the preset unsupervised feature extraction model is shown in the following formula (1): （1） in, represents the loss function of the preset unsupervised feature extraction model, represents the weight matrix of the preset unsupervised feature extraction model, is a constant, represents any of the initial sample speech acoustic features, represents the sigmoid function, represents the L1 norm, represents the transpose of the weight matrix, represents the L2 norm.

5. A speaker recognition device, characterized in that: The device comprises: A collection module, used to collect the target voice of the user; A recognition module, configured to use a speaker recognition model to perform speaker recognition on the target speech and output an identifier of a user corresponding to the target speech; wherein the speaker recognition model is trained using the method according to claim 1 or 2; The execution module is used to execute the target operation corresponding to the user according to the identifier of the user.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein the program executes the method according to any one of claims 1 to 3 when executed.

7. An electronic device, characterized in that: The method comprises a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to execute the method according to any one of claims 1 to 3 through the computer program.

Images