A voiceprint feature extraction method for specific content voice segments

Abstract

This application provides a method for extracting voiceprint features from speech segments with specific content. The method includes: obtaining an acoustic spectrum feature segment through preprocessing; constructing a time-delay neural network module; constructing a residual time-delay neural network module based on the time-delay neural network module, a weighted activation mechanism, and a residual structure; constructing a residual attention time-delay neural network module based on the time-delay neural network module, the residual time-delay neural network module, and an attention pooling mechanism; and inputting the acoustic spectrum feature segment into the residual attention time-delay neural network module to obtain the voiceprint features of the speech segment with specific content. The voiceprint feature extraction method provided here extracts deep-level information from features at multiple scales and, combined with residual networks, weighted activation, and attention pooling mechanisms, can effectively extract voiceprint features from speech segments with specific content.

Description

Technical Field

[0001] This invention relates to the field of speaker recognition technology, and in particular to a method for extracting voiceprint features for specific content speech segments. Background Technology

[0002] In recent years, with the development of pattern recognition and artificial intelligence, speaker recognition technology has made great strides and has been applied more and more widely, becoming one of the research hotspots in speech recognition technology.

[0003] Voiceprint recognition technology, also known as speaker identification technology, has important applications in information security, public security and judicial systems, and military defense. Currently, voiceprint recognition has achieved excellent performance on many datasets. With the widespread use of the internet and mobile devices, the importance of identity verification has become particularly prominent. In this context, the use of voiceprint passwords can enhance the security and reliability of account access on top of existing identity verification technologies. In practical applications, voiceprint recognition technology often uses specific settings, and due to the simplicity and universality of random number strings, it has become the mainstream trend for speaker identification technology in password applications.

[0004] However, due to the issue of co-pronunciation of specific content (i.e., the phenomenon that the pronunciation of a certain vocal organ is affected by the vocal organs before and after it), commonly used voiceprint feature extraction techniques do not take into account specific content and the limitation on the length of specific content (short speech segments) to improve user experience, resulting in poor performance of speaker recognition systems based on specific content in practical applications.

[0005] Currently, audio segments containing specific content are often only 2-4 seconds long, and ordinary algorithms cannot fully utilize the effective information in the audio features. How to better extract the acoustic features of short audio segments containing specific content has become a research focus in this field. Summary of the Invention

[0006] The purpose of this application is to address the deficiencies in the existing technology.

[0007] This application provides a method for extracting voiceprint features for specific content speech segments. The method takes into account the transient nature of specific content speech segments and extracts effective information from speech features at multiple scales. The voiceprint feature extraction method of this application is efficient, accurate, and has strong real-time performance.

[0008] This application provides a method for extracting voiceprint features from speech segments containing specific content. The method includes: determining text containing specific content and corresponding speech; extracting acoustic spectrum features of the corresponding speech; performing speech recognition on the acoustic spectrum features; segmenting the acoustic spectrum features to obtain acoustic spectrum feature segments corresponding to the text containing specific content; and constructing a residual attention delay neural network module based on a time-delay neural network module, a residual time-delay neural network module, and an attention pooling mechanism to output voiceprint features of the speech segment containing specific content from the acoustic spectrum feature segments. The time-delay neural network module is used to perform a one-dimensional convolution operation on the input feature information over time to extract acoustic feature information. The residual time-delay neural network module is constructed based on the time-delay neural network module, a weighted activation mechanism, and a residual structure to extract multi-scale feature information from the input feature information.

[0009] In a feasible embodiment, the step of extracting the speech acoustic spectrum features of the corresponding speech, performing speech recognition on the speech acoustic spectrum features, and segmenting the speech acoustic spectrum features to obtain speech acoustic spectrum feature segments corresponding to the text containing specific content includes: extracting acoustic spectrum features based on the digit string reading speech of a test speaker or registered speaker; performing end-to-end speech recognition on the acoustic spectrum features to obtain corresponding digit string text labels, as well as start and end time labels of corresponding digit segments and silence segments; segmenting the speech acoustic spectrum features according to the start and end time labels of the corresponding digit segments and silence segments, and removing silence segments to obtain acoustic spectrum feature segments corresponding to digit segments.

[0010] In one feasible embodiment, the step of extracting acoustic spectrum features based on the digital string reading speech of the test speaker or registered speaker includes: obtaining Mel-frequency cepstral features (MFCC) or perceptual linear prediction features (PLP) based on the digital string reading speech of the test speaker or registered speaker, performing differential cepstral feature (Delta) analysis on the features, obtaining a 60-dimensional feature vector for each frame, and then extracting the acoustic spectrum features.

[0011] In one feasible embodiment, the time-delay neural network module is used to perform a one-dimensional convolution operation on the input feature information over a time length to extract acoustic feature information, including: performing one-dimensional convolution on a time scale using a convolution kernel of a certain size to extract feature information and achieve the fusion of features from different channels; performing activation calculation on the output feature information after convolution using the LeakyReLU (Leaky Rectified Linear Unit) activation function; and performing batch standardization on the output feature information after the LeakyReLU activation function calculation using the BatchNorm function, wherein the mean and standard deviation of the BatchNorm function are calculated from the mean and standard deviation of each dimension.

[0012] In one feasible embodiment, the time-delay neural network module is used to perform a one-dimensional convolution operation on the input feature information over time to extract acoustic feature information, including: transposing the input feature information using the Transpose function; performing convolution processing on the transposed information using the Convolution function; standardizing the convolution-processed information using the LeakyReLU activation function and the BatchNorm function; and transposing and outputting the standardized information using the Transpose function.

[0013] In a feasible embodiment, the LeakyReLU activation function includes: LeakyReLU(x) = max(0,x) + αmin(0,x), where α is a small positive number; the BatchNorm function includes: Where gamma and beta are learnable parameter vectors.

[0014] In a feasible embodiment, the residual time-delay neural network module is constructed based on the time-delay neural network module, the weighted activation mechanism, and the residual structure. It is used to extract multi-scale feature information from input feature information, including: using at least one layer of the time-delay neural network module with different convolutional kernel sizes and strides; performing dimensionality reduction using a Linear1 linear function, activation using a ReLU activation function, dimensionality increase using a Linear2 linear function, and activation using a sigmoid activation function to obtain weighted activation coefficients; applying the weighted activation coefficients to the output feature information of the last time-delay neural network module to output the output feature information after the weighted activation mechanism; and adding the output feature information after the weighted activation mechanism to the input feature information of the first layer of the time-delay neural network module to obtain the output feature information of this layer.

[0015] In a feasible embodiment, the residual time-delay neural network module is constructed based on the time-delay neural network module, the weighted activation mechanism, and the residual structure, and is used to extract multi-scale feature information from the input feature information. It includes: multiple time-delay neural network modules, the weighted activation mechanism consisting of multiple linear layers and corresponding activation functions, and the residual structure extracting residual information and superimposing it with the original features (the input feature information of the first layer time-delay neural network module). In this process, grouped convolution is used to reduce the amount of computation.

[0016] In one feasible embodiment, the residual time-delay neural network module is constructed based on the time-delay neural network module, the weighted activation mechanism, and the residual structure, and is used to extract multi-scale feature information from the input feature information. Specifically, this module includes: multiple time-delay neural network modules that extract information from acoustic features through one-dimensional convolution operations over time; an introduction of a weighted activation mechanism to select the effective parts of the feature map; and the introduction of a residual structure to improve the performance of deep networks.

[0017] In a feasible embodiment, the ReLU activation function includes: ReLU(x) = max(0,x); the sigmoid activation function includes:

[0018] In a feasible embodiment, constructing a residual attention delay neural network module based on the delay neural network module, the residual delay neural network module, and the attention pooling mechanism includes: extracting feature information through at least one layer of delay neural network module and at least one layer of residual delay neural network module; performing one-dimensional convolution on the output feature information of the last layer of delay neural network module using the Convolution function, activating it with the tanh activation function, performing another one-dimensional convolution using the Convolution function, and then obtaining attention coefficients after passing through the softmax activation function; applying the attention coefficients to the output feature information of the last layer of delay neural network module to obtain the mean and standard deviation; concatenating the mean and standard deviation to output the feature information after the attention pooling mechanism; standardizing the feature information after the attention pooling mechanism using the BatchNorm function, and outputting the voiceprint features of a specific content speech segment through a linear layer.

[0019] In one feasible embodiment, the tanh activation function includes: The softmax activation function includes:

[0020] In a feasible embodiment, the construction of the residual attention delay neural network module includes: connecting the multi-layer delay neural network module and the multi-layer residual delay neural network module; the attention pooling mechanism consists of multiple one-dimensional convolutions and corresponding activation functions, and after obtaining the attention coefficients, they are applied to the input of the attention pooling mechanism to obtain the mean and standard deviation; batch standardization is performed using the BatchNorm function; finally, the confidence of the classification result is obtained through a linear layer.

[0021] This application provides a method for extracting voiceprint features from speech segments with specific content. The method includes: obtaining acoustic spectrum feature segments through a preprocessing stage; constructing a time-delay neural network module; constructing a residual time-delay neural network module based on the time-delay neural network module, a weighted excitation mechanism, and a residual structure; constructing a residual attention time-delay neural network module based on the time-delay neural network module, the residual time-delay neural network module, and an attention pooling mechanism; and inputting the acoustic spectrum feature segments into the residual attention time-delay neural network module to obtain the voiceprint features of the speech segment with specific content. The voiceprint feature extraction method provided here extracts deep-level information from features at multiple scales and, combined with residual networks, weighted excitation, and attention pooling mechanisms, can effectively model and recognize speech segments with specific content to obtain voiceprint features.

[0022] In the framework of the voiceprint feature extraction algorithm of this application, different network modules call each other. The voiceprint feature extraction technology of this application can perform refined and in-depth processing of the audio features of short, specific content speech segments, effectively improving the efficiency and accuracy of speaker representation of specific content speech. Commonly used voiceprint extraction techniques do not consider the speech content. This application provides a technique for extracting voiceprint features for speech segments with specific content, which can expand the application scenarios of voiceprint technology. Conventional neural network algorithms extract feature information from a certain angle, and it is difficult to extract rich voiceprint information for speech segments with specific content. The algorithm provided in this application extracts deep information of features from multiple scales and combines residual networks, weighted activation, attention mechanisms and other methods to effectively model and recognize speech segments with specific content. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating a method for extracting voiceprint features from a specific content speech segment according to an embodiment of the present invention.

[0024] Figure 2 This is a structural diagram of the time-delay neural network module according to an embodiment of the present invention;

[0025] Figure 3 This is a structural diagram of the residual delay neural network module according to an embodiment of the present invention;

[0026] Figure 4 This is a structural diagram of the residual attention delay neural network module in an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0028] To provide the public with a better understanding of this invention, certain specific details are described in detail below. However, those skilled in the art will fully understand this invention even without these detailed descriptions.

[0029] Figure 1 This is a flowchart illustrating a method for extracting voiceprint features from specific content speech segments according to an embodiment of the present invention. Figure 1 As shown, this application provides a method for extracting voiceprint features from specific content speech segments. The voiceprint feature extraction method includes the following steps:

[0030] Step S110: Determine the text containing specific content and its corresponding audio.

[0031] Step S120: Extract the acoustic spectrum features of the corresponding speech, perform speech recognition on the acoustic spectrum features, segment the acoustic spectrum features, and obtain the acoustic spectrum feature segments corresponding to the text containing specific content.

[0032] Step S130: Construct a time-delay neural network module to perform a one-dimensional convolution operation on the input feature information over time and output acoustic feature information.

[0033] Step S140: Based on the time-delay neural network module, the weighted activation mechanism, and the residual structure, construct a residual time-delay neural network module to output multi-scale feature information from the input feature information;

[0034] Step S150: Based on the time delay neural network module, the residual time delay neural network module, and the attention pooling mechanism, construct a residual attention time delay neural network module to output the voiceprint features of a specific content speech segment from the acoustic spectrum feature segment;

[0035] The acoustic spectrum feature fragments are input into the residual attention delay neural network module to obtain the voiceprint features of specific content speech fragments.

[0036] In step S110, text with specific content is determined based on the specific use scenario, and the text with specific content is read aloud by the trainer or tester to obtain speech containing the specific content. Step S110 is the step of determining the training speech or test speech for the voiceprint feature extraction method for speech segments with specific content.

[0037] In step S120, steps S110 and S120 are preprocessing stages for the training and test speech. Text readings of specific content are used to obtain speech containing that specific content, and the acoustic spectral features of this speech are extracted. Further, the extracted acoustic spectral features can be general Mel-frequency cepstral features (MFCC) or perceptual linear prediction features (PLP), and differential cepstral features (Delta) are applied to these features. Each frame yields a 60-dimensional feature vector, and the specific feature dimension can be selected according to the specific task scenario.

[0038] The acoustic spectrum features are subjected to speech recognition, and the acoustic spectrum features are segmented to obtain acoustic spectrum feature segments corresponding to the text containing specific content. Further, end-to-end speech recognition technology is used to perform speech recognition on the obtained acoustic spectrum features, and corresponding text tags for specific content, as well as start and end time tags for specific content segments and silent segments; based on the start and end time tags of the corresponding digit segments and silent segments, the speech acoustic spectrum features are segmented, and silent segments are removed to obtain acoustic spectrum feature segments.

[0039] In step S130, Figure 2 This is a structural diagram of the time-delay neural network module in an embodiment of the present invention, as shown below. Figure 2 As shown, the time-delay neural network module further specifically includes:

[0040] The Transpose function is used to transpose the input feature information.

[0041] The Convolution function is used to perform convolution processing on the transposed information to extract feature information and achieve the fusion of features from different channels.

[0042] The information after convolution is standardized using the LeakyReLU activation function and the BatchNorm function. LeakyReLU(x) = max(0,x) + αmin(0,x), where α is a small positive number. The BatchNorm function includes: Where gamma and beta are learnable parameter vectors. These are the mean and standard deviation of each dimension, which accelerates the convergence speed and improves the stability of the model.

[0043] The Transpose function transposes the standardized information and outputs the feature information.

[0044] In step S140, Figure 3 This is a structural diagram of the residual delay neural network module in an embodiment of the present invention, as shown below. Figure 3 As shown, step S140 includes: a three-layer time-delay neural network with different convolution kernel sizes and convolution strides constructed in step S130, and two linear layers. The two linear layers add weighted activation coefficients to the output feature information of the last time-delay neural network, extract residual information, and superimpose it with the original input signal. In this process, grouped convolution is used to reduce the amount of computation.

[0045] like Figure 3 As shown, the three-layer time-delay neural network constructed in step S130 of step S140 specifically includes:

[0046] The first-layer time-delay neural network module uses a convolution kernel of a certain size and a convolution stride to perform convolution, and extracts the input feature information in a precise manner;

[0047] The second-layer time-delay neural network module uses the output feature information of the first-layer time-delay neural network module as the input feature information, increases the size of the convolution kernel and the convolution stride, increases the receptive field, and extracts global information.

[0048] The third-layer time-delay neural network module uses the output feature information of the second-layer time-delay neural network module as the input feature information, and adopts a convolution kernel of size 1 to change the number of feature channels to achieve feature fusion of different channels;

[0049] like Figure 3 As shown, the area within the dashed box represents the weighted activation mechanism. This mechanism includes two linear layers (Linear1 function and Linear2 function) and two activation functions (ReLU activation function and sigmoid activation function), specifically:

[0050] The first linear layer uses the Linear1 linear function to reduce the dimensionality of the output feature information of the third-layer time-delay neural network module, and uses the ReLU activation function: ReLU(x) = max(0,x) for activation;

[0051] The second linear layer uses the Linear2 linear function to increase the dimensionality of the output features of the first linear layer, and uses the sigmoid activation function: Activation is performed to obtain the weighted incentive coefficients.

[0052] like Figure 3 As shown, after the feature information is processed by the weighted incentive mechanism, the weighted incentive coefficients are obtained. The weighted incentive coefficients are then applied to the input feature information of the first linear layer of the weighted incentive mechanism to weight different parts of the feature map of the input feature information, thereby achieving the optimization of feature information.

[0053] The input feature information of the first linear layer after weighting the excitation coefficients is superimposed with the input feature information of the first time-delay neural network module to obtain the output feature information of the residual time-delay neural network.

[0054] In step S150, a residual attention delay neural network module is constructed based on the time delay neural network module in step S120, the residual time delay neural network module in step S130, and the attention pooling mechanism. Figure 4 This is a structural diagram of the residual attention delay neural network module according to an embodiment of the present invention, as shown below. Figure 4 As shown, step S150 specifically includes: the four-layer time-delay neural network in step S120 and the three-layer residual time-delay neural network in step S130; the attention pooling mechanism consists of two one-dimensional convolutions and corresponding activation functions to obtain the attention coefficients of the features, which are applied to the input feature information of the attention pooling mechanism to obtain the mean and standard deviation; batch standardization is performed using the BatchNorm function; finally, the confidence of the classification result is obtained through the linear layer of the Linear function, thereby obtaining the voiceprint features of the specific content speech segment.

[0055] The time-delay neural network in the four-layer step S120 and the residual time-delay neural network in the three-layer step S130 further include:

[0056] The first-layer time-delay neural network module uses a certain size convolution kernel and convolution stride to extract local information of the input features of this layer network;

[0057] The second-layer time-delay neural network module uses the output feature information of the first-layer time-delay neural network module as the input feature information, and performs convolution using a convolution kernel of size 1 to change the number of feature channels and achieve feature fusion of different channels;

[0058] The third-layer residual delay neural network module uses the output feature information of the second-layer delay neural network module as the input feature information and performs convolution using a convolution kernel of a certain size;

[0059] The fourth residual delay neural network module uses the output feature information of the third residual delay neural network module as the input feature information, thereby increasing the size of the convolution kernel;

[0060] The fifth residual delay neural network module uses the output feature information of the fourth residual delay neural network module as the input feature information, thereby increasing the size of the convolution kernel;

[0061] The sixth-layer time delay neural network module uses the output feature information of the fifth-layer residual time delay neural network module as the input feature information, and the convolution kernel size is 1, integrating the feature information of each channel;

[0062] The seventh-layer time-delay neural network module uses the output feature information of the sixth-layer time-delay neural network module as the input feature information, and the convolution kernel size is 1, increasing the final required feature dimension.

[0063] The attention pooling mechanism consists of two one-dimensional convolutions and corresponding activation functions to obtain attention coefficients for the features. After being applied to the input feature information of the attention pooling mechanism, the mean and standard deviation are calculated, and the mechanism further includes:

[0064] In the attention pooling mechanism, the output feature information of the seventh-layer time-delay neural network module is subjected to one-dimensional convolution using the Convolution function, followed by a tanh activation function; after further convolution, it is then subjected to a softmax activation function to obtain attention coefficients; these attention coefficients are applied to the output feature information of the seventh-layer time-delay neural network module, and the mean and standard deviation are calculated. These mean and standard deviations are then concatenated together to serve as the output feature information of the attention pooling mechanism. The tanh activation function is as follows: The softmax activation function is

[0065] The BatchNorm function is used to batch standardize the output feature information of the attention pooling mechanism; finally, the confidence of the classification result is obtained through a linear layer of the Linear function, thereby obtaining the voiceprint features of a specific content speech segment.

[0066] This embodiment constructs a residual attention delay neural network module, the network structure of which is as follows: Figure 4 As shown, feature information is extracted sequentially through a two-layer time-delay neural network module, a three-layer residual time-delay neural network module, and a two-layer time-delay neural network module. Then, a one-dimensional convolution is performed using the Convolution function, followed by a tanh activation function and another one-dimensional convolution using the Convolution function. After passing through a softmax activation function, attention coefficients are obtained. These attention coefficients are applied to the output of the final time-delay neural network module, and their mean and standard deviation are calculated. The concatenated values ​​are then standardized using BatchNorm. A linear layer then extracts the voiceprint features of specific speech segments. Finally, matching these features with a template determines whether the speaker belongs to the same person.

[0067] The residual attention delay neural network module constructed in this embodiment is the main structure of a method for extracting voiceprint features from specific content speech segments. In actual use, only steps S110, S120, and S150 need to be explicitly used; the sub-network structures generated in steps S130 and S140 are already included in step S150.

[0068] The speaker feature extraction method for specific content speech segments in this embodiment also includes a training phase and a testing phase during use.

[0069] During the training phase, the trainable parameters in this algorithm are learned using the training data, and the features of the training data are used as the basis for step S150. Figure 4 The network structure shown takes input feature information and outputs output feature information representing the voiceprint features of a specific content speech segment. This output feature information is then matched with a template to determine whether the short speech segment and the training template belong to the same person. Alternatively, multi-objective training can be constructed to simultaneously train content recognition and voiceprint feature extraction, thereby effectively extracting voiceprint features for speech segments with specific content.

[0070] During the testing phase, according to the voiceprint feature extraction method of this embodiment, output feature information representing the voiceprint features of a specific content speech segment is obtained. Then, the output feature information is matched with the template to determine whether the short speech and the test template belong to the same person.

[0071] The time-delay neural network module in this embodiment is the basic structure of this algorithm and is invoked as a whole during use. The residual time-delay neural network module is the main structure for multi-scale feature extraction in this algorithm and is invoked as a whole during use. The residual attention time-delay neural network module is the main structure of a speaker feature extraction method for specific content speech segments and is invoked as a whole during use.

[0072] Regarding the "one-to-one" mode (determining whether the voice and the registration template are from the same person), the algorithm described in this embodiment is used to extract voiceprint features from specific content voice segments, and the voiceprint features are compared with the registration template to determine whether the voice and the registration template are from the same person.

[0073] Regarding the "one-to-many" mode (determining whether the speech belongs to the registered template library), the algorithm described in this embodiment is used to extract voiceprint features from specific content speech segments. These voiceprint features are then compared with the registered template library to determine whether the speech belongs to the library. Furthermore, if the usage scenario is a closed environment, speech not belonging to the registered template library is not processed; if the usage scenario is an open environment, speech not belonging to the registered template library is added to the library.

[0074] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. Therefore, it should be understood that the above description is only one specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for extracting voiceprint features from specific speech segments, characterized in that, The voiceprint feature extraction method includes: Identify text containing specific content and its corresponding audio; Extract the acoustic spectrum features of the corresponding speech, perform speech recognition on the acoustic spectrum features, segment the acoustic spectrum features, and obtain the acoustic spectrum feature fragments corresponding to the text containing specific content; Based on a time-delay neural network module, a residual time-delay neural network module, and an attention pooling mechanism, a residual attention time-delay neural network module is constructed to output the voiceprint features of a specific content speech segment from the acoustic spectrum feature segment. The time-delay neural network module is used to perform a one-dimensional convolution operation on the input feature information over time to extract acoustic feature information. The residual time-delay neural network module is constructed based on the time-delay neural network module, the weighted activation mechanism, and the residual structure to extract multi-scale feature information from the input feature information. The residual attention delay neural network module, constructed based on the time-delay neural network module, the residual time-delay neural network module, and the attention pooling mechanism, includes: Feature information is extracted using at least one layer of time-delay neural network module and at least one layer of residual time-delay neural network module; The output feature information of the last layer of the time-delay neural network module is convolved in one dimension using the Convolution function, activated by the tanh activation function, and then convolved in one dimension again using the Convolution function. After passing through the softmax activation function, the attention coefficient is obtained. The attention coefficient is applied to the output feature information of the last layer of the time-delay neural network module to obtain the mean and standard deviation. The mean and standard deviation are concatenated to obtain the output feature information of the attention pooling mechanism. The BatchNorm function is used to standardize the output feature information of the attention pooling mechanism, and the output voiceprint features of specific content speech segments are then passed through a linear layer.

2. The voiceprint feature extraction method according to claim 1, characterized in that, The process of extracting the acoustic spectral features of the corresponding speech, performing speech recognition on the acoustic spectral features, and segmenting the acoustic spectral features to obtain the acoustic spectral feature fragments corresponding to the text containing specific content includes: Acoustic spectral features are extracted from the digital string readings of test speakers or registered speakers. End-to-end speech recognition is performed on the acoustic spectrum features to obtain the corresponding digit string text labels, as well as the start and end time labels of the corresponding digit segments and silence segments; Based on the start and end time tags of the corresponding digital segment and silent segment, the acoustic spectrum features are segmented, and the silent segments are removed to obtain the acoustic spectrum feature segments corresponding to the digital segments.

3. The voiceprint feature extraction method according to claim 2, characterized in that, The extraction of acoustic spectral features from the digital string readings of the test speaker or registered speaker includes: Based on the digital string reading speech of the test speaker or registered speaker, Mel-Cepstral features or perceptual linear prediction features are obtained. Differential cepstral feature analysis is performed on these features, and a 60-dimensional feature vector can be obtained for each frame. Then, acoustic spectral features are extracted.

4. The voiceprint feature extraction method according to claim 1, characterized in that, The time-delay neural network module is used to perform a one-dimensional convolution operation on the input feature information over a time length to extract acoustic feature information, including: One-dimensional convolution is performed on the time scale using a convolution kernel of a certain size to extract feature information and achieve the fusion of features from different channels; The LeakyReLU activation function is used to calculate the activation of the output feature information after convolution. The output feature information calculated by the LeakyReLU activation function is batch standardized using the BatchNorm function. The mean and standard deviation of the BatchNorm function are calculated from the mean and standard deviation of each dimension.

5. The voiceprint feature extraction method according to claim 1, characterized in that, The residual time-delay neural network module is constructed based on the time-delay neural network module, the weighted activation mechanism, and the residual structure. It is used to extract multi-scale feature information from the input feature information, including: At least one layer of the aforementioned time-delay neural network module using different kernel sizes and strides; The dimensionality is reduced using the Linear1 linear function, activated by the ReLU activation function, increased by the Linear2 linear function, and activated by the sigmoid activation function to obtain the weighted activation coefficients. The weighted excitation coefficients are applied to the output feature information of the last time-delay neural network module, and the output feature information after the weighted excitation mechanism is output. The output feature information, which has undergone a weighted incentive mechanism, is added to the input feature information of the first-layer time-delay neural network module to obtain the output feature information of this layer of the network.

6. The voiceprint feature extraction method according to claim 5, characterized in that, The ReLU activation function includes: ; The sigmoid activation function includes: .

7. The voiceprint feature extraction method according to claim 1, characterized in that, The tanh activation function includes: .

Images