Voiceprint Recognition Model Training Method, Identity Recognition Method, Device and Electronic Equipment

Abstract

The embodiments of the present application provide a method for training a voiceprint recognition model, an identity recognition method, an apparatus, and an electronic device. The method for training a voiceprint recognition model includes: training a model to be trained using a training sample set to obtain a voiceprint recognition model; wherein, the model to be trained includes a plurality of first residual blocks connected in sequence, and each of the first residual blocks includes a first convolutional layer and a first skip connection; during the model training process, the specific processing method of each of the first residual blocks is: processing the input feature vector through the first convolutional layer and outputting the processed feature vector; fusing the feature vector transmitted to the output end of the first convolutional layer through the first skip connection and the feature vector output by the first convolutional layer, and outputting the fused feature vector. The embodiments of the present application improve the accuracy of the voiceprint recognition model in performing voiceprint recognition.

Description

Technical Field

[0001] This application belongs to the technical field of artificial intelligence, and particularly relates to a method for training a voiceprint recognition model, an identity recognition method, a device, and an electronic device. Background Art

[0002] With the development of artificial intelligence, more and more scenarios can perform identity recognition through artificial intelligence. Among various identity recognition means, voice is regarded as a promising feature for identifying personal identity and has been receiving increasing attention.

[0003] In related technologies, voice recognition is usually implemented by means of a neural network model. For example, voiceprint modeling is performed by extracting the attention features of speech signals, and a Gaussian Mixed Model (GMM) is used for feature mapping. However, this method is relatively primitive and has a low recognition accuracy. Summary of the Invention

[0004] The purpose of the embodiments of this application is to provide a method for training a voiceprint recognition model, an identity recognition method, a device, and an electronic device, which are used to improve the recognition accuracy.

[0005] In a first aspect, the embodiments of this application provide a method for training a voiceprint recognition model, and the method includes:

[0006] Training a model to be trained by using a training sample set to obtain a voiceprint recognition model;

[0007] Wherein, the model to be trained includes a plurality of first residual blocks connected in sequence, and each first residual block includes a first convolutional layer and a first skip connection;

[0008] During the model training process, the specific processing method of each first residual block is as follows:

[0009] Processing the input feature vector through the first convolutional layer and outputting the processed feature vector; fusing the feature vector transmitted to the output end of the first convolutional layer through the first skip connection and the feature vector output by the first convolutional layer, and outputting the fused feature vector.

[0010] It can be seen that in the embodiments of the present application, since each residual block included in the model to be trained contains a first convolutional layer and a skip connection, and the skip connection is a skip connection connecting the input end and the output end of the first convolutional layer, that is, a skip connection is implemented for each convolutional layer. Due to the addition of this shortcut of the skip connection, the gradient from the deep layer can directly and unobstructedly pass to the upper layer, so that during the process of training the basic model, it is less likely to have gradient dispersion, effectively alleviating the gradient dispersion phenomenon that is likely to occur when the number of convolutional layers is too large, improving the network training ability, and further improving the accuracy of the trained voiceprint recognition model, and ultimately improving the accuracy of voiceprint recognition using this voiceprint recognition model.

[0011] In a second aspect, an embodiment of the present application provides an identity recognition method, and the method includes:

[0012] Obtain the voiceprint feature data of the user to be recognized;

[0013] Input the voiceprint feature data into the voiceprint recognition model for feature extraction, and output the voice feature vector of the user to be recognized;

[0014] Perform identity recognition according to the voice feature vector;

[0015] Wherein, the voiceprint recognition model includes a plurality of second residual blocks connected in sequence, and each of the second residual blocks only includes a second convolutional layer;

[0016] During the process of using the voiceprint recognition model for feature extraction, the specific processing method of each of the second residual blocks is: processing the input feature vector through the second convolutional layer, and outputting the processed feature vector.

[0017] It can be seen that in the embodiments of the present application, each second residual block included in the voiceprint recognition model only includes a second convolutional layer. During the process of inputting the voiceprint feature data into the voiceprint recognition model for feature extraction, the input feature vector is processed through the second convolutional layer, and the processed feature vector is output, and identity recognition is performed through the output feature vector. Since the voiceprint recognition model is trained from the model to be trained, the accuracy of the voiceprint recognition model is guaranteed, thereby improving the accuracy of voiceprint recognition using this voiceprint recognition model; in addition, since the residual block included in the voiceprint recognition model only includes a second convolutional layer, compared with the model in the training stage, the skip connection is eliminated during the inference process, improving the inference efficiency of the feature extraction process.

[0018] In a third aspect, an embodiment of the present application provides a voiceprint recognition model training device, and the device includes:

[0019] A training unit, configured to train the model to be trained using a training sample set to obtain a voiceprint recognition model;

[0020] Among them, the model to be trained includes a plurality of first residual blocks connected in sequence, and each of the first residual blocks includes a first convolutional layer and a first skip connection;

[0021] During the model training process, the specific processing method of each of the first residual blocks is as follows:

[0022] The input feature vector is processed through the first convolutional layer, and the processed feature vector is output; the feature vector transmitted to the output end of the first convolutional layer through the first skip connection and the feature vector output by the first convolutional layer are fused, and the fused feature vector is output.

[0023] In a fourth aspect, an embodiment of the present application provides an identity recognition device, which includes:

[0024] An acquisition unit, configured to acquire voiceprint feature data of a user to be recognized;

[0025] A feature extraction unit, configured to input the voiceprint feature data into a voiceprint recognition model for feature extraction, and output a voice feature vector of the user to be recognized;

[0026] An identity recognition unit, configured to perform identity recognition according to the voice feature vector;

[0027] Among them, the voiceprint recognition model includes a plurality of second residual blocks connected in sequence, and each of the second residual blocks only includes a second convolutional layer;

[0028] During the process of using the voiceprint recognition model for feature extraction, the specific processing method of each of the second residual blocks is as follows: the input feature vector is processed through the second convolutional layer, and the processed feature vector is output.

[0029] In a fifth aspect, an embodiment of the present application provides an electronic device, including a processor, a memory, and a program or instruction stored on the memory and executable on the processor. When the program or instruction is executed by the processor, the steps of the method described in the first aspect or the second aspect are implemented.

[0030] In a sixth aspect, an embodiment of the present application provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the method described in the first aspect or the second aspect are implemented.

[0031] In a seventh aspect, an embodiment of the present application provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor, and the processor is configured to run a program or instruction to implement the method described in the first aspect or the second aspect. Description of the Drawings

[0032] To more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments recorded in the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0033] Figure 1 It is a schematic flow chart of the voiceprint recognition model training method in the embodiments of the present application;

[0034] Figure 2 It is a schematic diagram of the first residual block in the embodiments of the present application;

[0035] Figure 3 It is a schematic diagram of the second residual block in the embodiments of the present application;

[0036] Figure 4 It is a schematic diagram of parameter reconstruction for the third residual block in the embodiments of the present application;

[0037] Figure 5 It is a step flow chart of the identity recognition method in the embodiments of the present application;

[0038] Figure 6 It is a schematic diagram of the voiceprint recognition model in the embodiments of the present application;

[0039] Figure 7 It is a schematic diagram of the module composition of the voiceprint recognition model training device in the embodiments of the present application;

[0040] Figure 8 It is a schematic diagram of the module composition of the identity recognition device in the embodiments of the present application;

[0041] Figure 9 It is a schematic diagram of the structure of the electronic device in the embodiments of the present application. Detailed implementation manners

[0042] The following will clearly describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art belong to the scope of protection of the present application.

[0043] The terms "first," "second," and the like in the specification and claims of this application are used to distinguish similar objects, and are not used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "first," "second," and the like generally distinguish objects of a class and do not limit the number of objects. For example, the first object may be one or more.

[0044] Furthermore, the terms "include," "comprise," and "have," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements, but may optionally include steps or elements not listed, or may optionally include other steps or elements inherent to the process, method, product, or apparatus.

[0045] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0046] Currently, in consumer finance, the flow of funds between companies and customers during the loan process is substantial, creating the potential for fraud during customer service interactions. Furthermore, numerous methods for forging user identities are already available, and facial recognition technology is easily compromised. Given this reality, voice is increasingly considered a promising method for identifying individuals. Theoretically, a voiceprint is like a fingerprint; rarely do two people have identical voiceprint characteristics. The uniqueness of a voiceprint is primarily determined by two factors: the dimensions of the vocal cavity, including the throat, nasal cavity, and oral cavity; and the manipulation of the vocal organs, which include the lips, teeth, tongue, soft palate, and palatal muscles. The way these organs work together is learned through interaction with those around us, gradually forming our own unique characteristics.

[0047] The technology of using voiceprints to distinguish different people has been widely recognized and applied in various fields. It also has strong application prospects in the field of financial fraud prevention. Related technologies usually use network training models to perform identity recognition based on voiceprints, but the accuracy rate is relatively low.

[0048] To solve the above problems, the embodiments of the present application provide a method for training a voiceprint recognition model and an identity recognition method. First, a training sample set is used to train a model to be trained, and a voiceprint recognition model is obtained. The model to be trained includes a plurality of first residual blocks connected in sequence, and each first residual block includes a first convolutional layer and a first skip connection, which provides a basis for subsequent parameter reconstruction to obtain the voiceprint recognition model. In addition, the voiceprint recognition model includes a plurality of second residual blocks connected in sequence, and each second residual block only includes a second convolutional layer. In the process of inputting voiceprint feature data into the voiceprint recognition model for feature extraction, the input feature vector is processed by the second convolutional layer, and the processed feature vector is output. Since the voiceprint recognition model is trained from the model to be trained, the accuracy of the voiceprint recognition model is guaranteed, thereby improving the accuracy of voiceprint recognition using the voiceprint recognition model; and, since the residual blocks included in the voiceprint recognition model only include the second convolutional layer, compared with the model in the training stage, the skip connection is eliminated during the inference process, and the inference efficiency of the feature extraction process is improved.

[0049] Next, with reference to the accompanying drawings, the method for training a voiceprint recognition model and the identity recognition method provided by the embodiments of the present application will be described in detail through specific embodiments and their application scenarios.

[0050] As Figure 1 shown, it is a flowchart of the steps of the method for training a voiceprint recognition model provided by the embodiments of the present application, and the method includes the following steps:

[0051] Step 101: Use a training sample set to train a model to be trained to obtain a voiceprint recognition model.

[0052] Among them, the model to be trained includes a plurality of first residual blocks connected in sequence, and each first residual block includes a first convolutional layer and a first skip connection.

[0053] During the model training process, the specific processing method of each first residual block is as follows:

[0054] The input feature vector is processed through the first convolutional layer, and the processed feature vector is output; the feature vector transmitted to the output end of the first convolutional layer through the first skip connection and the feature vector output by the first convolutional layer are fused, and the fused feature vector is output.

[0055] Specifically, the training sample set includes voiceprint features labeled with user identity tags. Among them, when obtaining the training sample set, the voiceprint features can be extracted from voice files labeled with user identity tags.

[0056] In addition, specifically, the model to be trained includes a plurality of first residual blocks connected in sequence. Each first residual block includes a first convolutional layer and a first skip connection, where the first skip connection is a skip connection connecting the input end and the output end of the first convolutional layer.

[0057] It can be understood that the specific number of the first residual blocks is not specifically limited here, and the specific number of the first residual blocks can be set according to the actual situation.

[0058] During the training process of the model to be trained, the first convolutional layer can process the input feature vector and output the processed feature vector. The first skip connection can directly transmit the input feature vector to the output end of the first convolutional layer. Then, the residual block fuses the feature vector transmitted to the output end of the first convolutional layer through the first skip connection and the feature vector output by the first convolutional layer through an addition operation, and outputs the fused feature vector. Specifically, the fused feature vector can be represented by the following formula:

[0059] H(x) = F(x) + x;

[0060] Among them, H(x) represents the fused feature vector, x represents the input feature vector, and F(x) represents the processed feature vector output by the first convolutional layer after processing the input feature vector.

[0061] As an example, here, the case where the number of the first residual blocks is 2 is used to illustrate the first residual blocks included in the model to be trained. For example, as Figure 2 shown, in Figure 2 there are two residual blocks, namely the first residual block A1 and the first residual block A2. In each first residual block, there is a first convolutional layer 21 and a first skip connection 22, and the size of the first convolutional layer 21 is 3×3×64. Among them, taking the first residual block A1 as an example, the first convolutional layer 21 in the first residual block A1 can process the input feature vector and output the processed feature vector. The first skip connection 22 in the first residual block A1 outputs the input feature vector, and then the first residual block A1 fuses the feature vector transmitted to the output end of the first convolutional layer through the first skip connection and the feature vector output by the first convolutional layer through an addition operation, and outputs the fused feature vector.

[0062] In this way, since each residual block included in the model to be trained contains the first convolutional layer and the skip connection, the skip connection is a skip connection connecting the input and output ends of the first convolutional layer. Due to the addition of the skip connection shortcut, the gradient from the deep layer can pass directly to the previous layer without obstruction, so that in the process of training the basic model, gradient diffusion is less likely to occur, which effectively alleviates the gradient diffusion phenomenon that is prone to occur when there are too many convolutional layers, improves the network training ability, and then improves the accuracy of the trained voiceprint recognition model, and ultimately improves the accuracy of voiceprint recognition using the voiceprint recognition model; in addition, each first residual block includes the first convolutional layer and the first skip connection, which provides a basis for subsequent parameter reconstruction of the model to obtain a voiceprint recognition model.

[0063] In one possible implementation, the voiceprint recognition model includes multiple second residual blocks connected in sequence, and each second residual block includes only a second convolutional layer; in the process of feature extraction using the voiceprint recognition model, the specific processing method of each second residual block is: processing the input feature vector through the second convolutional layer, and outputting the processed feature vector.

[0064] Specifically, the voiceprint recognition model includes multiple sequentially connected second residual blocks, each of which includes only a second convolutional layer and no skip connections. When using this voiceprint recognition model for feature extraction, the second convolutional layer processes the input feature vector and outputs the processed feature vector.

[0065] As an example, following the above example, the trained voiceprint recognition model includes the second residual block as follows Figure 3 As shown. Figure 3 , including two residual blocks, a second residual block B1 and a second residual block B2. Each second residual block includes a second convolutional layer 31, and the size of the second convolutional layer 31 is 3×3×64. Taking the second residual block B1 as an example, the second convolutional layer 31 in the second residual block B1 can process the input feature vector and output the processed feature vector.

[0066] Compared with the model in the training stage, the voiceprint recognition model gets rid of skip connections during the reasoning process, which improves the reasoning efficiency of the feature extraction process.

[0067] In one possible implementation, using the training sample set to train the model to be trained to obtain the voiceprint recognition model may include the following steps:

[0068] The to-be-trained model is trained using a training sample set to obtain a trained model; wherein, the trained model includes a plurality of sequentially connected third residual blocks, and each of the third residual blocks includes a third convolutional layer and a second skip connection; parameter reconstruction is performed on the third residual blocks included in the trained model to obtain the voiceprint recognition model.

[0069] Specifically, in the process of obtaining the voiceprint recognition model, the to-be-trained model can be first trained using a training sample set to obtain a trained model, and then parameter reconstruction is performed on the third residual blocks included in the trained model to obtain the voiceprint recognition model.

[0070] Among them, in the process of using the trained model for feature extraction, the specific processing method of each third residual block is as follows: the input feature vector is processed through the third convolutional layer, and the processed feature vector is output; the feature vector transmitted to the output end of the third convolutional layer through the second skip connection and the feature vector output by the third convolutional layer are fused, and the fused feature vector is output.

[0071] It can be understood that after the parameter reconstruction of the third residual block, the output of the residual block after parameter reconstruction does not change compared with the output of the third residual block before parameter reconstruction, but only the parameters of the residual block change, that is, the feature vector output by the residual block after parameter reconstruction (i.e., the second residual block) is the fused feature vector output by the third residual block before parameter reconstruction.

[0072] Optionally, the second skip connection is a skip connection connecting the output end and the input end of the third convolutional layer; the parameter reconstruction of the third residual blocks included in the trained model to obtain the voiceprint recognition model may include the following steps:

[0073] A fourth convolutional layer is added to each of the third residual blocks; wherein, the input end of the fourth convolutional layer is connected to the input end of the third convolutional layer, the output end of the fourth convolutional layer is connected to the output end of the third convolutional layer, the weights of the other parameters except the normalization parameters in the fourth convolutional layer are all 0, the bias of the fourth convolutional layer is 0, and the convolutional size of the fourth convolutional layer is the same as that of the third convolutional layer;

[0074] The weights of the third convolutional layer and the fourth convolutional layer are added together, and the biases of the third convolutional layer and the fourth convolutional layer are added together to obtain the voiceprint recognition model.

[0075] Specifically, adding a fourth convolutional layer to each third residual block can also be understood as inserting the fourth convolutional layer into the channel formed by the second skip connection, and specific reference can be made to Figure 4 .

[0076] Among them, inFigure 4 In it, the third residual block C includes a third convolutional layer 41 and a second skip connection 42. After adding a fourth convolutional layer 43 to the third residual block C, since the weights of the other parameters in the fourth convolutional layer 43 except the normalization parameters are all 0, the bias of the fourth convolutional layer 43 is 0, and the convolutional size of the fourth convolutional layer 43 is the same as that of the third convolutional layer 41, so that after adding the weights of the third convolutional layer 41 and the fourth convolutional layer 43 and adding the biases of the third convolutional layer 41 and the fourth convolutional layer 43, the second convolutional layer 31 after parameter reconstruction can be obtained, that is, the second residual block is obtained, and thus the voiceprint recognition model is obtained.

[0077] In this way, through the above method, the parameter reconstruction of the third residual block included in the trained model is realized, so that the obtained voiceprint recognition model completely gets rid of the skip connection in the inference process, and the inference efficiency is improved.

[0078] In addition, optionally, the above first convolutional layer, second convolutional layer and third convolutional layer are all 3×3 convolutional layers. In this way, the superposition of two 3×3 convolutions can add one more non-linear transformation, thereby enhancing the model ability.

[0079] In the voiceprint recognition model training method provided in this embodiment, since each residual block included in the model to be trained contains a first convolutional layer and a skip connection, and the skip connection is a skip connection connecting the input end and the output end of the first convolutional layer, due to the addition of this shortcut of the skip connection, the gradient from the deep layer can directly and unobstructedly pass to the upper layer, so that in the process of training the basic model, it is less likely to appear gradient dispersion, effectively alleviating the gradient dispersion phenomenon that is likely to occur when the number of convolutional layers is too large, improving the network training ability, and then improving the accuracy of the trained voiceprint recognition model, and finally improving the accuracy of voiceprint recognition using the voiceprint recognition model; in addition, each first residual block includes a first convolutional layer and a first skip connection, which provides a basis for subsequent parameter reconstruction of the model to obtain a voiceprint recognition model.

[0080] In addition, as Figure 5 shown, it is a step flow chart of the identity recognition method in the embodiment of the present application, and this method includes the following steps:

[0081] Step 501: Obtain the voiceprint feature data of the user to be recognized.

[0082] Specifically, when obtaining the voiceprint feature data of the user to be recognized, the voice data of the user to be recognized can be obtained, and the voiceprint feature data can be extracted from the voice data. For example, in the scenario of a financial customer service communicating with a user, when the user calls in, the conversation recording between the customer service and the user can be obtained and the voice channels can be separated, the user's voice channel can be separated, and the user's voiceprint feature data can be extracted.

[0083] The voiceprint feature data may include Fbank features. Fbank features are features widely used in voiceprint recognition, speech emotion recognition, speech recognition, and speech synthesis. In the process of obtaining Fbank features, the audio signal can be pre-emphasized, framed, and windowed first, and then the short-time Fourier transform (STFT) is performed on each frame of the signal to obtain the short-time amplitude spectrum. Finally, the Fbank features are obtained through the Mel filter bank from the short-time amplitude spectrum.

[0084] It should be noted that when extracting Fbank features, toolkits such as kaldi, librosa, or espnet can be used for feature extraction, or feature extraction can also be directly performed using Python. Specific limitations are not imposed here.

[0085] Step 502: Input the voiceprint feature data into the voiceprint recognition model for feature extraction, and output the voice feature vector of the user to be recognized.

[0086] Among them, the voiceprint recognition model includes a plurality of second residual blocks connected in sequence, and each of the second residual blocks only includes a second convolutional layer;

[0087] In the process of using the voiceprint recognition model for feature extraction, the specific processing method of each of the second residual blocks is: processing the input feature vector through the second convolutional layer and outputting the processed feature vector.

[0088] The following describes the extraction process of the above voice feature vector in combination with a specific voiceprint recognition model.

[0089] As Figure 6 shown, the voiceprint recognition model includes an input layer, a normalization convolutional layer, a fully connected layer, 4 second residual blocks (second residual block 1, second residual block 2, second residual block 3, and second residual block 4), 2 shallow feature fusion network layers (temp1 and temp2), and a splicing layer. The input layer, the normalization convolutional layer, the 4 second residual blocks, and the fully connected layer are connected in series in sequence. Specifically, the residual block can be represented as ResNetBlock.

[0090] Among them, the normalized convolutional layer is used to process the feature vectors input by the input layer, and the processed feature vectors are input into the second residual block 1. The second residual block 1 is used to process the processed feature vectors and output the first feature vector. The second residual block 2 is used to process the first feature vector and output the second feature vector. The second residual block 3 is used to process the input second feature vector and output the third feature vector. The second residual block 4 is used to process the input third feature vector and output the fourth feature vector. The shallow feature fusion network layer (temp1) is used to fuse the third feature vector P5 output by the second residual block 3 with the second feature vector output by the second residual block 2 to obtain the feature vector P4. The shallow feature fusion network layer (temp2) is used to fuse the first feature vector output by the second residual block 1 with the feature vector P4 to obtain the feature vector P3. The concatenation layer is used to fuse the feature vectors P3, P4 and P5, the output vector of the second residual block 4 and the output vector of the fully connected layer to obtain the voiceprint feature vector.

[0091] Step 503: Perform identity recognition based on the voice feature vector.

[0092] Specifically, when performing identity recognition based on the voice feature vector, the voice feature vector and the pre-obtained registered voice vectors can be input into a preset classification model to output the identity recognition result.

[0093] Specifically, the preset classification model can include an aamsoftmax classifier or a Probabilistic Linear Discriminant Analysis (PLDA) classification model, etc., and specific limitations are not imposed here.

[0094] In addition, the registered voice vectors can be the voice vectors pre-registered by multiple users, and the registered voice vectors can represent the identities of the users. For example, in the scenario of a financial customer service agent talking to a customer, the user corresponding to the registered voice vector can be a user whose identity has been verified. Subsequently, a registered voice can be recorded using the user's voice, and the registered voice vector can be extracted.

[0095] Of course, the registered voice vectors can also be obtained by feature extraction through a voiceprint recognition model, so as to ensure that the presentation form of the registered voice vectors is the same as that of the voiceprint feature data of the user to be recognized, facilitating the classification and comparison of the voiceprint feature data and the registered voice vectors.

[0096] Thus, in this embodiment, since the voiceprint recognition model is obtained by training the model to be trained, the accuracy of the voiceprint recognition model is ensured, thereby improving the accuracy of voiceprint recognition using the voiceprint recognition model. In addition, since the residual block included in the voiceprint recognition model only includes the second convolutional layer and has no skip connection, only the input feature vector is processed through the second convolutional layer, and the processed feature vector is output. Compared with the model in the training stage, this enables the skip connection to be eliminated during the inference process, improving the inference efficiency of the feature extraction process.

[0097] It should be noted that for the voiceprint recognition model training method provided in the embodiments of the present application, the execution subject may be a voiceprint recognition model training device, or a control module in the voiceprint recognition model training device for executing the voiceprint recognition model training method. In the embodiments of the present application, the voiceprint recognition model training method is executed by the voiceprint recognition model training device as an example to illustrate the voiceprint recognition model training device provided in the embodiments of the present application.

[0098] As Figure 7 shown, the voiceprint recognition model training device includes a training unit 701; wherein,

[0099] The training unit 701 is configured to train the model to be trained using a training sample set to obtain a voiceprint recognition model;

[0100] Wherein, the model to be trained includes a plurality of first residual blocks connected in sequence, and each of the first residual blocks includes a first convolutional layer and a first skip connection;

[0101] During the model training process, the specific processing method of each of the first residual blocks is as follows:

[0102] The input feature vector is processed through the first convolutional layer, and the processed feature vector is output; the feature vector transmitted to the output end of the first convolutional layer through the first skip connection and the feature vector output by the first convolutional layer are fused, and the fused feature vector is output.

[0103] In a possible implementation manner, the voiceprint recognition model includes a plurality of second residual blocks connected in sequence, and each of the second residual blocks only includes a second convolutional layer;

[0104] During the process of using the voiceprint recognition model for feature extraction, the specific processing method of each of the second residual blocks is as follows: the input feature vector is processed through the second convolutional layer, and the processed feature vector is output.

[0105] In a possible implementation, the training unit 701 is configured to train a model to be trained using a training sample set to obtain a trained model; wherein, the trained model includes a plurality of third residual blocks connected in sequence, and each of the third residual blocks includes a third convolutional layer and a second skip connection;

[0106] Perform parameter reconstruction on the third residual blocks included in the trained model to obtain the voiceprint recognition model.

[0107] In a possible implementation, the second skip connection is a skip connection connecting the output end and the input end of the third convolutional layer; the training unit 701 is configured to add a fourth convolutional layer to each of the third residual blocks; wherein, the input end of the fourth convolutional layer is connected to the input end of the third convolutional layer, the output end of the fourth convolutional layer is connected to the output end of the third convolutional layer, the weights of the remaining parameters in the fourth convolutional layer except the normalization parameters are all 0, the bias of the fourth convolutional layer is 0, and the convolutional size of the fourth convolutional layer is the same as that of the third convolutional layer; add the weights of the third convolutional layer and the fourth convolutional layer, and add the biases of the third convolutional layer and the fourth convolutional layer to obtain the voiceprint recognition model.

[0108] In a possible implementation, the first convolutional layer, the second convolutional layer, and the third convolutional layer are all 3×3 convolutional layers.

[0109] The voiceprint recognition model training device provided by the embodiments of the present application can implement Figures 1 to 4 Each process implemented by the voiceprint recognition model training method in the method embodiment described above will not be repeated here to avoid repetition.

[0110] In addition, for the identity recognition method provided by the embodiments of the present application, the execution subject may be an identity recognition device, or a control module in the identity recognition device for executing the identity recognition method. In the embodiments of the present application, the identity recognition device executing the identity recognition method is taken as an example to illustrate the identity recognition device provided by the embodiments of the present application.

[0111] As Figure 8 shown, the identity recognition device includes an acquisition unit 801, a feature extraction unit 802, and an identity recognition unit 803; wherein,

[0112] The acquisition unit 801 is configured to acquire voiceprint feature data of a user to be recognized;

[0113] The feature extraction unit 802 is configured to input the voiceprint feature data into a voiceprint recognition model for feature extraction, and output a voice feature vector of the user to be recognized;

[0114] An identity recognition unit 803, configured to perform identity recognition according to the voice feature vector;

[0115] Wherein, the voiceprint recognition model includes a plurality of second residual blocks connected in sequence, and each of the second residual blocks only includes a second convolutional layer;

[0116] During the process of using the voiceprint recognition model for feature extraction, the specific processing method of each of the second residual blocks is: processing the input feature vector through the second convolutional layer, and outputting the processed feature vector.

[0117] The identity recognition device provided by the embodiments of the present application can implement Figure 5 and Figure 6 each process implemented by the identity recognition method in the method embodiments. To avoid repetition, details are not described herein again.

[0118] The identity recognition device or the voiceprint recognition model training device in the embodiments of the present application may be a device, or a component, an integrated circuit, or a chip in a terminal. The device may be a mobile electronic device or a non-mobile electronic device. Exemplarily, the mobile electronic device may be a mobile phone, a tablet computer, a laptop computer, a handheld computer, a vehicle-mounted electronic device, a wearable device, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (PDA), etc., and the non-mobile electronic device may be a server, a Network Attached Storage (NAS), a personal computer (PC), a television (TV), a teller machine, or a self-service machine, etc. The embodiments of the present application do not make specific limitations.

[0119] The identity recognition device or the voiceprint recognition model training device in the embodiments of the present application may be a device with an operating system. The operating system may be an Android operating system, an iOS operating system, or other possible operating systems. The embodiments of the present application do not make specific limitations.

[0120] Based on the same technical concept, the embodiments of the present application further provide an electronic device, which is used to execute the above-mentioned voiceprint recognition model training method or identity recognition method, Figure 9A structural schematic diagram of an electronic device for implementing various embodiments of the present application. The electronic device may vary greatly due to configuration or performance differences, and may include a processor 910, a communications interface 920, a memory 930, and a communication bus 940. Among them, the processor 910, the communications interface 920, and the memory 930 complete mutual communication through the communication bus 940. The processor 910 can call a computer program stored in the memory 930 and executable on the processor 910 to perform the following steps:

[0121] Use the training sample set to train the model to be trained to obtain a voiceprint recognition model;

[0122] Among them, the model to be trained includes a plurality of first residual blocks connected in sequence, and each of the first residual blocks includes a first convolutional layer and a first skip connection;

[0123] During the model training process, the specific processing method of each of the first residual blocks is as follows:

[0124] Process the input feature vector through the first convolutional layer and output the processed feature vector; fuse the feature vector transmitted to the output end of the first convolutional layer through the first skip connection and the feature vector output by the first convolutional layer, and output the fused feature vector.

[0125] Alternatively, the processor 910 is used to perform the following steps:

[0126] Obtain the voiceprint feature data of the user to be identified;

[0127] Input the voiceprint feature data into the voiceprint recognition model for feature extraction, and output the voice feature vector of the user to be identified;

[0128] Perform identity recognition based on the voice feature vector;

[0129] Among them, the voiceprint recognition model includes a plurality of second residual blocks connected in sequence, and each of the second residual blocks only includes a second convolutional layer;

[0130] During the process of using the voiceprint recognition model for feature extraction, the specific processing method of each of the second residual blocks is as follows: process the input feature vector through the second convolutional layer and output the processed feature vector.

[0131] The embodiments of the present application further provide a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, each process of the above-mentioned identity recognition method or the voiceprint recognition model training method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be elaborated here.

[0132] Wherein, the processor is the processor in the electronic device described in the above embodiment. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk or optical disc, etc.

[0133] The embodiments of the present application further provide a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run a program or instruction to implement each process of the above method embodiment, and the same technical effect can be achieved. To avoid repetition, it will not be elaborated here.

[0134] It should be understood that the chip mentioned in the embodiments of the present application may also be referred to as a system-on-chip, system chip, chip system, or system-on-chip, etc.

[0135] It should be noted that in this article, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising one..." does not exclude the existence of additional identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in a reverse order according to the functions involved. For example, the described methods may be performed in an order different from that described, and various steps may be added, omitted, or combined. Additionally, the features described with reference to certain examples may be combined in other examples.

[0136] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-described example methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented through hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions for causing a terminal (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in various embodiments of the present application.

[0137] The embodiments of the present application have been described above in conjunction with the accompanying drawings. However, the present application is not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative and not restrictive. Under the inspiration of the present application, those of ordinary skill in the art can also make many forms without departing from the purpose of the present application and the scope protected by the claims, and all of them belong to the protection scope of the present application.

Claims

1. A method for training a voiceprint recognition model, characterized in that The method includes: Training a model to be trained using a training sample set to obtain a voiceprint recognition model; Wherein, the model to be trained includes a plurality of first residual blocks connected in sequence, and each of the first residual blocks includes a first convolutional layer and a first skip connection; During the model training process, the specific processing method of each of the first residual blocks is: Processing the input feature vector through the first convolutional layer and outputting the processed feature vector; fusing the feature vector transmitted to the output end of the first convolutional layer through the first skip connection and the feature vector output by the first convolutional layer, and outputting the fused feature vector; The voiceprint recognition model includes a plurality of second residual blocks connected in sequence, and each of the second residual blocks only includes a second convolutional layer, and the voiceprint recognition model is obtained by parameter reconstruction of the trained model to be trained.

2. The method according to claim 1, characterized in that During the process of using the voiceprint recognition model for feature extraction, the specific processing method of each of the second residual blocks is: processing the input feature vector through the second convolutional layer and outputting the processed feature vector.

3. The method according to claim 2, wherein The training of the model to be trained using the training sample set to obtain a voiceprint recognition model includes: Training the model to be trained using the training sample set to obtain a trained model; wherein, the trained model includes a plurality of third residual blocks connected in sequence, and each of the third residual blocks includes a third convolutional layer and a second skip connection; Performing parameter reconstruction on the third residual blocks included in the trained model to obtain the voiceprint recognition model.

4. The method according to claim or 3, characterized in that, The second skip connection is a skip connection connecting the output end and the input end of the third convolutional layer; the performing parameter reconstruction on the third residual blocks included in the trained model to obtain the voiceprint recognition model includes: Adding a fourth convolutional layer to each of the third residual blocks; wherein, the input end of the fourth convolutional layer is connected to the input end of the third convolutional layer, the output end of the fourth convolutional layer is connected to the output end of the third convolutional layer, the weights of the other parameters except the normalization parameters in the fourth convolutional layer are all 0, the bias of the fourth convolutional layer is 0, and the convolutional size of the fourth convolutional layer is the same as that of the third convolutional layer; Adding the weights of the third convolutional layer and the fourth convolutional layer, and adding the biases of the third convolutional layer and the fourth convolutional layer to obtain the voiceprint recognition model.

5. The method according to claim 4, characterized in that, The first convolutional layer, the second convolutional layer, and the third convolutional layer are all 3×3 convolutional layers.

6. An identity recognition method, characterized in that, The method includes: Obtaining the voiceprint feature data of the user to be recognized; Inputting the voiceprint feature data into the voiceprint recognition model for feature extraction, and outputting the voice feature vector of the user to be recognized; Performing identity recognition according to the voice feature vector; Wherein, the voiceprint recognition model includes a plurality of second residual blocks connected in sequence, and each of the second residual blocks only includes a second convolutional layer; During the process of using the voiceprint recognition model for feature extraction, the specific processing method of each of the second residual blocks is: processing the input feature vector through the second convolutional layer and outputting the processed feature vector.

7. An acoustic fingerprint recognition model training device, characterized in that The device includes: A training unit for training a model to be trained using a training sample set to obtain a voiceprint recognition model; Wherein, the model to be trained includes a plurality of first residual blocks connected in sequence, and each of the first residual blocks includes a first convolutional layer and a first skip connection; During the model training process, the specific processing method of each of the first residual blocks is as follows: Processing the input feature vector through the first convolutional layer and outputting the processed feature vector; fusing the feature vector transmitted to the output end of the first convolutional layer through the first skip connection and the feature vector output by the first convolutional layer, and outputting the fused feature vector; The voiceprint recognition model includes a plurality of second residual blocks connected in sequence, and each of the second residual blocks only includes a second convolutional layer, and the voiceprint recognition model is obtained by parameter reconstruction of the trained model to be trained.

8. An identity recognition device, characterized in that, The device includes: An acquisition unit for acquiring the voiceprint feature data of the user to be recognized; A feature extraction unit for inputting the voiceprint feature data into the voiceprint recognition model for feature extraction and outputting the voice feature vector of the user to be recognized; An identity recognition unit for performing identity recognition according to the voice feature vector; Wherein, the voiceprint recognition model includes a plurality of second residual blocks connected in sequence, and each of the second residual blocks only includes a second convolutional layer; During the process of using the voiceprint recognition model for feature extraction, the specific processing method of each of the second residual blocks is as follows: processing the input feature vector through the second convolutional layer and outputting the processed feature vector.

9. An electronic device, characterized in that, It includes a processor, a memory, and a program or instruction stored on the memory and executable on the processor. When the program or instruction is executed by the processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A readable storage medium, characterized in that, A program or instruction is stored on the readable storage medium. When the program or instruction is executed by the processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images