A voiceprint recognition method based on neural network and related device

Abstract

The present application discloses a voiceprint recognition method based on a neural network and a related device. The method includes: constructing a semi-orthogonal decomposition neural network model based on multiple semi-orthogonal convolution blocks, each semi-orthogonal convolution block includes multiple semi-orthogonal one-dimensional convolution layers, and the semi-orthogonal one-dimensional convolution layers are connected by serial connection, inner skip connection structure and outer skip connection structure; performing voiceprint recognition training on the semi-orthogonal decomposition neural network model according to a preset MFCC training set corresponding to the training voiceprint information to obtain a target recognition network model; using the target recognition network model to recognize the target voiceprint to obtain a voiceprint recognition result. Multiple semi-orthogonal one-dimensional convolution layers can decompose the original parameter matrix in the network, compress the redundant parameter expression space, filter out noise interference and reduce the delay span. The present application can solve the technical problems that the existing voiceprint recognition technology has poor noise resistance and limited delay modeling capability, resulting in a lack of accuracy and reliability in the recognition results.

Description

Technical Field

[0001] The present application relates to the field of voiceprint recognition technology, and in particular to a voiceprint recognition method and related devices based on a neural network. Background Art

[0002] Voiceprint recognition systems automatically identify speakers based on the characteristics of their voices. Bodyprint recognition, a type of biometric verification technology, verifies a speaker's identity through voice. This technology offers advantages such as convenience, stability, scalability, and security, and is commonly used in banking, social security, public security, smart homes, and mobile payments.

[0003] Existing voiceprint recognition methods are limited by the influence of noise in voiceprint information, which makes the recognition results lack accuracy and reliability. The time delay modeling capability of voiceprint recognition based on neural network models is limited, which leads to poor actual voiceprint recognition results and cannot meet high-standard application requirements. Summary of the Invention

[0004] The present application provides a neural network-based voiceprint recognition method and related devices, which are used to solve the technical problems that existing voiceprint recognition technologies have poor noise resistance and limited time delay modeling capabilities, resulting in a lack of accuracy and reliability in recognition results.

[0005] In view of this, the first aspect of the present application provides a voiceprint recognition method based on a neural network, comprising:

[0006] A semi-orthogonal decomposition neural network model is constructed based on multiple semi-orthogonal convolution blocks, each of which includes multiple semi-orthogonal one-dimensional convolution layers, and the semi-orthogonal one-dimensional convolution layers are connected by series connection, inner skip connection structure and outer skip connection structure;

[0007] Performing voiceprint recognition training on the semi-orthogonal decomposition neural network model according to a preset MFCC training set corresponding to the training voiceprint information to obtain a target recognition network model;

[0008] The target recognition network model is used to recognize the target voiceprint to obtain a voiceprint recognition result.

[0009] Preferably, the voiceprint recognition training is performed on the semi-orthogonal decomposition neural network model according to the preset MFCC training set corresponding to the training voiceprint information to obtain the target recognition network model, and the method further includes:

[0010] Performing a preprocessing operation on the training voiceprint information to obtain an audio frame to be processed, wherein the preprocessing operation includes emphasis, framing and windowing processing;

[0011] Based on the Fourier transform algorithm, the Mel filter is used to calculate the audio frame to obtain MFCC features;

[0012] A preset MFCC training set is constructed according to the MFCC features.

[0013] Preferably, the method of performing voiceprint recognition training on the semi-orthogonal decomposition neural network model according to the preset MFCC training set corresponding to the training voiceprint information to obtain the target recognition network model further includes:

[0014] A semi-orthogonal decomposition feature extractor is constructed based on multiple semi-orthogonal convolution blocks;

[0015] Perform voiceprint feature extraction training on the semi-orthogonal decomposition feature extractor according to a preset MFCC training set corresponding to the training voiceprint information to obtain a target voiceprint feature extractor;

[0016] During the voiceprint information registration process, the target voiceprint feature extractor is used to extract features from the newly added voiceprint, and the extracted voiceprint features are stored in a database.

[0017] Preferably, the voiceprint recognition training is performed on the semi-orthogonal decomposition neural network model according to the preset MFCC training set corresponding to the training voiceprint information to obtain the target recognition network model, and then the method further includes:

[0018] Performing a voiceprint recognition test on the target recognition network model using a preset MFCC test set corresponding to the test voiceprint information to obtain a test result;

[0019] Screening the target recognition network model according to the test results to obtain an optimized recognition network model;

[0020] Accordingly, the target recognition network model is used to recognize the target voiceprint to obtain the voiceprint recognition result, including:

[0021] The optimized recognition network model is used to recognize the target voiceprint to obtain a voiceprint recognition result.

[0022] The second aspect of the present application provides a voiceprint recognition device based on a neural network, comprising:

[0023] A model construction module is used to construct a semi-orthogonal decomposition neural network model based on multiple semi-orthogonal convolution blocks, each of which includes multiple semi-orthogonal one-dimensional convolution layers, and the semi-orthogonal one-dimensional convolution layers are connected by series connection, inner skip connection structure and outer skip connection structure;

[0024] A model training module is used to perform voiceprint recognition training on the semi-orthogonal decomposition neural network model according to a preset MFCC training set corresponding to the training voiceprint information to obtain a target recognition network model;

[0025] The voiceprint recognition module is used to use the target recognition network model to recognize the target voiceprint and obtain a voiceprint recognition result.

[0026] Preferably, it also includes:

[0027] A preprocessing module is used to perform preprocessing operations on the training voiceprint information to obtain audio frames to be processed. The preprocessing operations include emphasis, framing and windowing.

[0028] A feature extraction module is used to calculate the audio frame using a Mel filter based on a Fourier transform algorithm to obtain MFCC features;

[0029] The training set construction module is used to construct a preset MFCC training set according to the MFCC features.

[0030] Preferably, it also includes:

[0031] An extractor building module for building a semi-orthogonal decomposition feature extractor based on multiple semi-orthogonal convolution blocks;

[0032] An extractor training module is used to perform voiceprint feature extraction training on the semi-orthogonal decomposition feature extractor according to a preset MFCC training set corresponding to the training voiceprint information to obtain a target voiceprint feature extractor;

[0033] The extractor using module is used to extract features of the newly added voiceprint through the target voiceprint feature extractor during the voiceprint information registration process, and save the extracted voiceprint features in the database.

[0034] Preferably, it also includes:

[0035] A testing module, configured to perform a voiceprint recognition test on the target recognition network model using a preset MFCC test set corresponding to the test voiceprint information to obtain a test result;

[0036] An optimization module, configured to screen the target recognition network model according to the test results to obtain an optimized recognition network model;

[0037] Accordingly, the voiceprint recognition module is specifically used to:

[0038] The optimized recognition network model is used to recognize the target voiceprint to obtain a voiceprint recognition result.

[0039] A third aspect of the present application provides a voiceprint recognition device based on a neural network, the device comprising a processor and a memory;

[0040] The memory is used to store program code and transmit the program code to the processor;

[0041] The processor is used to execute the neural network-based voiceprint recognition method described in the first aspect according to the instructions in the program code.

[0042] In a fourth aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store program code, and the program code is used to execute the neural network-based voiceprint recognition method described in the first aspect.

[0043] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0044] In the present application, a voiceprint recognition method based on a neural network is provided, comprising: constructing a semi-orthogonal decomposition neural network model based on multiple semi-orthogonal convolution blocks, each semi-orthogonal convolution block comprising multiple semi-orthogonal one-dimensional convolution layers, and the semi-orthogonal one-dimensional convolution layers being connected in series, by an inner skip connection structure and an outer skip connection structure; performing voiceprint recognition training on the semi-orthogonal decomposition neural network model according to a preset MFCC training set corresponding to the training voiceprint information to obtain a target recognition network model; and using the target recognition network model to identify the target voiceprint to obtain a voiceprint recognition result.

[0045] The neural network-based voiceprint recognition method provided by this application uses a skip connection structure to connect the convolutional layers in the process of constructing a semi-orthogonal decomposition neural network model, directly transmitting the shallow voiceprint feature information to the deep convolutional layer, allowing the deep network to obtain richer voiceprint information, thereby improving the network's noise resistance. Multiple semi-orthogonal one-dimensional convolutional layers can decompose the original parameter matrix in the network, compressing the redundant parameter expression space, filtering out noise interference, and reducing the delay span, thereby achieving the purpose of long-delay learning. Therefore, this application can solve the technical problems of existing voiceprint recognition technology, which has poor noise resistance and limited delay modeling capabilities, resulting in a lack of accuracy and reliability in recognition results. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A flowchart of a neural network-based voiceprint recognition method provided in an embodiment of the present application;

[0047] Figure 2 A schematic diagram of the structure of a voiceprint recognition device based on a neural network provided in an embodiment of the present application;

[0048] Figure 3 A schematic diagram of the structure of a semi-orthogonal decomposition neural network model provided in an embodiment of the present application;

[0049] Figure 4 Schematic diagram of the network structure of the semi-orthogonal convolution block provided in an embodiment of the present application. DETAILED DESCRIPTION

[0050] In order to help those skilled in the art better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0051] For easier understanding, see Figure 1 , an embodiment of a voiceprint recognition method based on a neural network provided by this application includes:

[0052] Step 101: construct a semi-orthogonal decomposition neural network model based on multiple semi-orthogonal convolution blocks, where each semi-orthogonal convolution block includes multiple semi-orthogonal one-dimensional convolution layers, and the semi-orthogonal one-dimensional convolution layers are connected by series connection, inner skip connection structure and outer skip connection structure.

[0053] It should be noted that the current mainstream deep neural network (Delay Neural Network, DNN) and voiceprint recognition applications are based on the Time Delay Neural Network (TDNN), and the penultimate or second hidden layer of TDNN is used as the voiceprint feature output, called x-vector. TDNN is mainly built with multi-layer one-dimensional convolutional neural network (CNN) components. However, the one-dimensional convolution component has the ability to describe image or speech features at multiple scales, which is better than the ordinary fully connected method. However, TDNN's recognition effect decreases in strong noise environments and its noise resistance is insufficient. In addition, TDNN's time delay modeling capability is limited, and it can only effectively learn within a short and stable time range.

[0054] Therefore, this embodiment uses a semi-orthogonal convolutional layer to decompose the neural network's weight matrix, significantly reducing the number of parameters in the original one-dimensional convolutional weight layer. In supervised learning of speaker labels, the semi-orthogonal decomposition neural network, based on limited parameters, extracts important speaker voiceprint information in each factorization and filters out irrelevant noise information, thereby demonstrating noise immunity. Furthermore, the semi-orthogonal convolutional layer has a limited delay modeling range. If the span is set too large, missed samples will occur, impairing information filtering. Therefore, this embodiment uses multiple semi-orthogonal convolutional blocks, enabling learning over a large delay range with convolutional blocks of smaller spans, enhancing contextual learning capabilities while simultaneously combating noise.

[0055] See also Figure 3, the semi-orthogonal decomposition neural network model in this example includes multiple unequal semi-orthogonal convolution blocks, and adjacent semi-orthogonal convolution blocks are connected in series. In addition, starting from the second semi-orthogonal convolution block of the network model, each semi-orthogonal convolution block will splice one, two or more outer skip connection structures, or the output of the inner skip connection structure. It should be noted that the outer skip connection structure also includes network structures such as semi-orthogonal one-dimensional convolution layers and activation functions, and its output is transmitted to the second semi-orthogonal convolution block and the subsequent deep semi-orthogonal convolution blocks in the network model. The inner skip connection structure starts from the second semi-orthogonal convolution block, and the shallow feature information needs to be transmitted to the subsequent deep semi-orthogonal convolution blocks. From Figure 3 As can be seen from the example given, the input of the second semi-orthogonal convolution block includes a serial input and an outer skip connection input; the input of the third semi-orthogonal convolution block includes a serial input, an outer skip connection input, and an inner skip connection input, and so on. Specifically, the number of semi-orthogonal convolution blocks included in the network model can be determined according to actual needs and is not limited here, as long as it conforms to the convolution block connection concept of this embodiment.

[0056] The semi-orthogonal convolution block includes multiple semi-orthogonal one-dimensional convolution layers, and the number is at least 2, so it is a multi-segment semi-orthogonal convolution block, see Figure 4 , Figure 4 The curved arrows in the figure represent the data flow of the inner and outer skip connection structures. In addition to the semi-orthogonal one-dimensional convolutional layer, the semi-orthogonal convolution block also includes an outer skip splicing layer, an activation function, a regularization layer, and an output layer. The output layer includes the input information of the inner skip connection structure and the output information of the regularization layer. The outer skip splicing layer can receive the output information of the outer skip connection structure input by other semi-orthogonal convolution blocks and combine it with other information received by the layer in a splicing manner. It can be understood that the outer skip splicing layer can receive the output information of multiple outer skip connection structures of different sizes.

[0057] It should be noted that the semi-orthogonal one-dimensional convolutional layer can decompose the input parameter matrix A[a,a], and constrain the decomposition parameter matrix M[a,b] to conform to the semi-orthogonal decomposition, so that the effective voiceprint information of the output matrix B[b,b] can be preserved. Among them, a and b are the matrix dimensions. The formula for constrained decomposition is as follows:

[0058]

[0059] A=MB

[0060] Where α is a floating-point coefficient, which defaults to 1; I is the identity matrix.

[0061] When the constraints converge, we have:

[0062] A=MB≈MM T

[0063] That is, the output matrix B is approximately equal to the parameter matrix M T .

[0064] The parameter matrix N[b,a] of the one-dimensional convolution performs a convolution operation on the decomposed matrix B or the concatenated matrix. Multiple convolution kernels learn information at different scales to generate the matrix P[a,a]. Among them, the matrix P is different from the output result R[a,a] of the ordinary one-dimensional convolution TDNN. Assuming that the parameter matrix of the ordinary one-dimensional convolution TDNN is W[a,a], the specific differences between the ordinary one-dimensional convolution and the semi-orthogonal decomposition one-dimensional convolution are as follows:

[0065] Ordinary one-dimensional convolution: A*W→R

[0066] Semi-orthogonal decomposition of one-dimensional convolution: A*MN→P

[0067] When the dimensionality reduction of the parameter matrix M is less than or equal to a / 4, the total number of parameters of M and N will be less than or equal to half the number of parameters of W. The multiple audio clips of each training speaker will have different noise factors. In supervised learning of speaker labels, the neural network will learn the common characteristics of speakers under various noise conditions. Semi-orthogonal one-dimensional convolution decomposes the original matrix and compresses the redundant parameter representation space, making it possible to refine speaker information and filter out noise interference. The amount of noise information modeled is:

[0068] ε=W-MN.

[0069] The activation function and the regularization layer perform nonlinear activation processing on the matrix P output by the semi-orthogonal one-dimensional convolutional layer and then regularize it to obtain the hidden layer information Q; finally, the output layer performs integration operations such as addition or splicing on the activated and regularized matrix Q and the jump connection input matrix A. This embodiment selects weighted addition, and the weight defaults to 0.66. After weight addition, an output matrix O is formed.

[0070] Step 102: Perform voiceprint recognition training on the semi-orthogonal decomposition neural network model according to the preset MFCC training set corresponding to the training voiceprint information to obtain a target recognition network model.

[0071] The data used to train the semi-orthogonal decomposition neural network model is acoustic features, namely Mel-Frequency Cepstrum Coefficients (MFCC). For labeled training sets, the model output and the labels can be used to calculate the cross-entropy loss, thereby optimizing the model training and obtaining the target recognition network model.

[0072] Furthermore, before step 102, the following steps are also included:

[0073] Perform preprocessing on the training voiceprint information to obtain the audio frame to be processed. The preprocessing operations include emphasis, framing and windowing.

[0074] Based on the Fourier transform algorithm, the Mel filter is used to calculate the audio frame to obtain the MFCC features;

[0075] Construct a preset MFCC training set based on MFCC features.

[0076] Furthermore, step 102 further includes:

[0077] A semi-orthogonal decomposition feature extractor is constructed based on multiple semi-orthogonal convolution blocks;

[0078] Perform voiceprint feature extraction training on the semi-orthogonal decomposition feature extractor according to the preset MFCC training set corresponding to the training voiceprint information to obtain the target voiceprint feature extractor;

[0079] During the voiceprint information registration process, the target voiceprint feature extractor is used to extract features from the newly added voiceprint, and the extracted voiceprint features are saved in the database.

[0080] The feature extractor shares the same network structure as the semi-orthogonal decomposition neural network model. Essentially, it outputs voiceprint features in the first layer of the final pooling layer of the semi-orthogonal decomposition neural network model, bypassing the final fully connected layer for outputting recognition results. The training process for the target voiceprint feature extractor is the same as that for the recognition model; the feature extractor can be used for both initial voice registration and validation set acquisition.

[0081] Furthermore, step 102 further includes:

[0082] Use the preset MFCC test set corresponding to the test voiceprint information to perform voiceprint recognition test on the target recognition network model to obtain the test results;

[0083] The target recognition network model is screened according to the test results to obtain the optimized recognition network model;

[0084] Accordingly, step 103 includes:

[0085] The target voiceprint is identified using an optimized recognition network model to obtain the voiceprint recognition result.

[0086] Step 103: Use the target recognition network model to identify the target voiceprint and obtain a voiceprint recognition result.

[0087] The neural network-based voiceprint recognition method provided in the embodiment of the present application uses a skip connection structure to connect the convolutional layers in the process of constructing a semi-orthogonal decomposition neural network model, directly transmitting the shallow voiceprint feature information to the deep convolutional layer, so that the deep network can obtain richer voiceprint information, thereby improving the network's noise resistance. Multiple semi-orthogonal one-dimensional convolutional layers can decompose the original parameter matrix in the network, compressing the redundant parameter expression space, filtering out noise interference, and reducing the delay span, thereby achieving the purpose of long-delay learning. Therefore, the embodiment of the present application can solve the technical problems of existing voiceprint recognition technology, such as poor noise resistance and limited delay modeling capabilities, which lead to a lack of accuracy and reliability in recognition results.

[0088] For easier understanding, see Figure 2 , the present application provides an embodiment of a voiceprint recognition device based on a neural network, comprising:

[0089] A model construction module 201 is used to construct a semi-orthogonal decomposition neural network model based on multiple semi-orthogonal convolution blocks, each of which includes multiple semi-orthogonal one-dimensional convolution layers, and the semi-orthogonal one-dimensional convolution layers are connected by series connection, inner skip connection structure and outer skip connection structure;

[0090] The model training module 202 is used to perform voiceprint recognition training on the semi-orthogonal decomposition neural network model according to the preset MFCC training set corresponding to the training voiceprint information to obtain a target recognition network model;

[0091] The voiceprint recognition module 203 is used to recognize the target voiceprint using a target recognition network model to obtain a voiceprint recognition result.

[0092] Furthermore, it also includes:

[0093] The pre-processing module 204 is used to perform pre-processing operations on the training voiceprint information to obtain audio frames to be processed. The pre-processing operations include emphasis, framing and windowing.

[0094] The feature extraction module 205 is used to calculate the audio frame using a Mel filter based on a Fourier transform algorithm to obtain MFCC features;

[0095] The training set construction module 206 is configured to construct a preset MFCC training set according to the MFCC features.

[0096] Furthermore, it also includes:

[0097] An extractor construction module 207, configured to construct a semi-orthogonal decomposition feature extractor based on a plurality of semi-orthogonal convolution blocks;

[0098] The extractor training module 208 is used to perform voiceprint feature extraction training on the semi-orthogonal decomposition feature extractor according to the preset MFCC training set corresponding to the training voiceprint information to obtain a target voiceprint feature extractor;

[0099] The extractor using module 209 is used to extract features of the newly added voiceprint through the target voiceprint feature extractor during the voiceprint information registration process, and store the extracted voiceprint features in the database.

[0100] Furthermore, it also includes:

[0101] The testing module 210 is used to perform a voiceprint recognition test on the target recognition network model using a preset MFCC test set corresponding to the test voiceprint information to obtain a test result;

[0102] The optimization module 211 is used to screen the target recognition network model according to the test results to obtain an optimized recognition network model;

[0103] Accordingly, the voiceprint recognition module 203 is specifically used to:

[0104] The target voiceprint is identified using an optimized recognition network model to obtain the voiceprint recognition result.

[0105] The present application also provides a voiceprint recognition device based on a neural network, the device including a processor and a memory;

[0106] The memory is used to store program codes and transmit the program codes to the processor;

[0107] The processor is used to execute the neural network-based voiceprint recognition method in the above method embodiment according to the instructions in the program code.

[0108] The present application also provides a computer-readable storage medium, which is used to store program code, and the program code is used to execute the neural network-based voiceprint recognition method in the above method embodiment.

[0109] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0110] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0111] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0112] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for executing all or part of the steps of the method described in each embodiment of the present application through a computer device (which can be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (English full name: Read-Only Memory, English abbreviation: ROM), random access memory (English full name: Random Access Memory, English abbreviation: RAM), disk or optical disk and other media that can store program code.

[0113] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A voiceprint recognition method based on neural network, characterized in that: include: A semi-orthogonal decomposition neural network model is constructed based on multiple semi-orthogonal convolution blocks, each of which includes multiple semi-orthogonal one-dimensional convolution layers, and the semi-orthogonal one-dimensional convolution layers are connected by series connection, inner skip connection structure and outer skip connection structure; The semi-orthogonal convolution block also includes an outer skip splicing layer, an activation function, a regularization layer and an output layer, wherein the output layer includes input information integrated into the inner skip connection structure and output information of the regularization layer; The outer jump splicing layer receives the output information of the outer jump connection structure input by other semi-orthogonal convolution blocks, and combines it with other information received by this layer in a splicing manner. The outer jump splicing layer receives the output information of multiple unequal outer jump connection structures; The semi-orthogonal decomposition neural network model includes a plurality of unequal semi-orthogonal convolution blocks, adjacent semi-orthogonal convolution blocks are connected in series, and starting from the second semi-orthogonal convolution block of the network model, each semi-orthogonal convolution block is spliced with one, two or more outputs of the outer skip connection structure or the inner skip connection structure; The outer skip connection structure includes a semi-orthogonal one-dimensional convolution layer and an activation function network structure, the output of which is transmitted to the second semi-orthogonal convolution block in the network model and the subsequent deep semi-orthogonal convolution block; The inner skip connection structure starts from the second semi-orthogonal convolution block, and the shallow feature information is transmitted to the subsequent deep semi-orthogonal convolution block; The input of the second semi-orthogonal convolution block includes a serial input and an outer skip connection input; the input of the third semi-orthogonal convolution block includes a serial input, an outer skip connection input and an inner skip connection input; Performing voiceprint recognition training on the semi-orthogonal decomposition neural network model according to a preset MFCC training set corresponding to the training voiceprint information to obtain a target recognition network model; The target recognition network model is used to recognize the target voiceprint to obtain a voiceprint recognition result.

2. The voiceprint recognition method based on neural network according to claim 1, characterized in that: The method further includes: performing voiceprint recognition training on the semi-orthogonal decomposition neural network model according to the preset MFCC training set corresponding to the training voiceprint information to obtain a target recognition network model; Performing a preprocessing operation on the training voiceprint information to obtain an audio frame to be processed, wherein the preprocessing operation includes emphasis, framing and windowing processing; Based on the Fourier transform algorithm, the Mel filter is used to calculate the audio frame to obtain MFCC features; A preset MFCC training set is constructed according to the MFCC features.

3. The voiceprint recognition method based on neural network according to claim 1, characterized in that: The method further comprises: performing voiceprint recognition training on the semi-orthogonal decomposition neural network model according to the preset MFCC training set corresponding to the training voiceprint information to obtain a target recognition network model; A semi-orthogonal decomposition feature extractor is constructed based on multiple semi-orthogonal convolution blocks; Perform voiceprint feature extraction training on the semi-orthogonal decomposition feature extractor according to a preset MFCC training set corresponding to the training voiceprint information to obtain a target voiceprint feature extractor; During the voiceprint information registration process, the target voiceprint feature extractor is used to extract features from the newly added voiceprint, and the extracted voiceprint features are stored in a database.

4. The voiceprint recognition method based on neural network according to claim 1, characterized in that: The method further includes performing voiceprint recognition training on the semi-orthogonal decomposition neural network model according to the preset MFCC training set corresponding to the training voiceprint information to obtain a target recognition network model, and then further including: Performing a voiceprint recognition test on the target recognition network model using a preset MFCC test set corresponding to the test voiceprint information to obtain a test result; Screening the target recognition network model according to the test results to obtain an optimized recognition network model; Accordingly, the target recognition network model is used to recognize the target voiceprint to obtain the voiceprint recognition result, including: The optimized recognition network model is used to recognize the target voiceprint to obtain a voiceprint recognition result.

5. A voiceprint recognition device based on neural network, characterized in that: include: A model construction module is used to construct a semi-orthogonal decomposition neural network model based on multiple semi-orthogonal convolution blocks, each of which includes multiple semi-orthogonal one-dimensional convolution layers, and the semi-orthogonal one-dimensional convolution layers are connected by series connection, inner skip connection structure and outer skip connection structure; The semi-orthogonal convolution block also includes an outer skip splicing layer, an activation function, a regularization layer and an output layer, wherein the output layer includes input information integrated into the inner skip connection structure and output information of the regularization layer; The outer jump splicing layer receives the output information of the outer jump connection structure input by other semi-orthogonal convolution blocks, and combines it with other information received by this layer in a splicing manner. The outer jump splicing layer receives the output information of multiple unequal outer jump connection structures; The semi-orthogonal decomposition neural network model includes a plurality of unequal semi-orthogonal convolution blocks, adjacent semi-orthogonal convolution blocks are connected in series, and starting from the second semi-orthogonal convolution block of the network model, each semi-orthogonal convolution block is spliced with one, two or more outputs of the outer skip connection structure or the inner skip connection structure; The outer skip connection structure includes a semi-orthogonal one-dimensional convolution layer and an activation function network structure, the output of which is transmitted to the second semi-orthogonal convolution block in the network model and the subsequent deep semi-orthogonal convolution block; The inner skip connection structure starts from the second semi-orthogonal convolution block, and the shallow feature information is transmitted to the subsequent deep semi-orthogonal convolution block; The input of the second semi-orthogonal convolution block includes a serial input and an outer skip connection input; the input of the third semi-orthogonal convolution block includes a serial input, an outer skip connection input and an inner skip connection input; A model training module is used to perform voiceprint recognition training on the semi-orthogonal decomposition neural network model according to a preset MFCC training set corresponding to the training voiceprint information to obtain a target recognition network model; The voiceprint recognition module is used to use the target recognition network model to recognize the target voiceprint and obtain a voiceprint recognition result.

6. The neural network-based voiceprint recognition device according to claim 5, characterized in that: Also includes: A preprocessing module is used to perform preprocessing operations on the training voiceprint information to obtain audio frames to be processed. The preprocessing operations include emphasis, framing and windowing. A feature extraction module is used to calculate the audio frame using a Mel filter based on a Fourier transform algorithm to obtain MFCC features; The training set construction module is used to construct a preset MFCC training set according to the MFCC features.

7. The voiceprint recognition device based on neural network according to claim 5, characterized in that: Also includes: An extractor building module for building a semi-orthogonal decomposition feature extractor based on multiple semi-orthogonal convolution blocks; An extractor training module is used to perform voiceprint feature extraction training on the semi-orthogonal decomposition feature extractor according to a preset MFCC training set corresponding to the training voiceprint information to obtain a target voiceprint feature extractor; The extractor using module is used to extract features of the newly added voiceprint through the target voiceprint feature extractor during the voiceprint information registration process, and save the extracted voiceprint features in the database.

8. The voiceprint recognition device based on neural network according to claim 5, characterized in that: Also includes: A testing module, configured to perform a voiceprint recognition test on the target recognition network model using a preset MFCC test set corresponding to the test voiceprint information to obtain a test result; An optimization module, configured to screen the target recognition network model according to the test results to obtain an optimized recognition network model; Accordingly, the voiceprint recognition module is specifically used to: The optimized recognition network model is used to recognize the target voiceprint to obtain a voiceprint recognition result.

9. A voiceprint recognition device based on a neural network, characterized in that: The device includes a processor and a memory; The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the neural network-based voiceprint recognition method according to any one of claims 1 to 4 according to the instructions in the program code.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium is used to store program code, and the program code is used to execute the neural network-based voiceprint recognition method according to any one of claims 1 to 4.

Images