A Voiceprint Recognition Method Based on DR-Res2net Module

Abstract

This invention provides a voiceprint recognition method based on the DR-Res2net module. This method provides rich and effective feature information when processing voiceprint data, exhibits strong generalization ability, and has a lower error rate in classification, thus achieving more ideal recognition results. In this invention, the characteristics of the dense DenseNet model are integrated into the Res2Net model to construct the DR-Res2net model. In the voice recognition model, DR-Res2Block is used as the core module for voiceprint data recognition. During the recognition process, each DR-Res2net module performs residual and dense connections on each output feature simultaneously to obtain richer features. The addition processing in the module increases the amount of information contained in each feature, while the concatenation processing ensures that the features include both high-semantic low-resolution and low-semantic high-resolution features, preserving features from different receptive fields to the greatest extent possible.

Description

Technical Field

[0001] This invention relates to the field of voiceprint recognition technology, specifically a voiceprint recognition method based on the DR-Res2net module. Background Technology

[0002] With the rapid development of science and technology, the internet has brought convenience to people's lives, but it has also led to frequent cases of account and password theft. Therefore, the demand for identity verification is increasing, and identity authentication has gradually become a hot topic of common concern. Traditional identification methods can no longer meet people's needs for personal information security, and biometrics has gradually become the focus of identity authentication technology. It is a technology that uses a person's personalized information to achieve identity verification. Commonly used biometric methods include facial recognition, voiceprint recognition, and fingerprint recognition. Compared with traditional identification methods, these biometric technologies have advantages such as strong anti-counterfeiting performance and less susceptibility to being forgotten or lost. Among them, voiceprint recognition is an important branch of the field of speech recognition.

[0003] In the current field of voiceprint recognition technology, voiceprint recognition models are typically built based on convolutional neural networks, such as LeNet, VGG, TDNN, or Res2Net, which are all commonly used voiceprint recognition models. Figure 1 The diagram shown illustrates an example of the connection relationships between feature layers in the Res2Block module of the Res2Net model. Res2Net is a convolutional neural network with residual connections. Figure 1 In the Res2Block shown, the features are first divided in dimensions to obtain feature layers x1 to x4. Then, feature layers x1 to x4 are fed into Conv, BN, and ReLU layers respectively. The features in the middle layer need to be added with the features processed by the previous layer before Conv, BN, and ReLU are performed. The last feature layer x4 does not need to be operated on and is directly used as z4 to synthesize the features z1 to z3 after all x1 to x3 are processed into the total feature.

[0004] In the Res2Net network, multi-scale features are represented at the granular level, and the receptive field of each network layer is increased. The feature maps learned by each layer in Res2Block are also directly passed to all subsequent layers as input. However, due to the insufficient generalization ability of the Res2Net model, some features are lost during the recognition process, resulting in a high error rate in classification and thus a less than ideal performance in recognizing voiceprint data. Summary of the Invention

[0005] To address the issues of high error rates and unsatisfactory recognition results in existing Res2Net-based voiceprint recognition methods, this invention provides a voiceprint recognition method based on the DR-Res2net module. This method provides rich and effective feature information when processing voiceprint data, has strong generalization ability, and exhibits a lower error rate in classification, thereby achieving more ideal recognition results.

[0006] The technical solution of this invention is as follows: a voiceprint recognition method based on the DR-Res2net module, comprising the following steps:

[0007] S1: Construct an acoustic model and train the acoustic model to obtain a trained acoustic model;

[0008] S2: Collect the voiceprint data to be processed, preprocess the original voiceprint signal, and output the voiceprint frame sequence;

[0009] S3: Perform time-domain and frequency-domain analysis on the voiceprint frame sequence, extract the Mel frequency cepstral coefficients, and output the feature vector sequence to be identified;

[0010] S4: After processing, the sequence of feature vectors to be identified is input into the trained acoustic model for classification and recognition. The resulting recognition result is the classification result of voiceprint recognition.

[0011] Its features are:

[0012] The acoustic model includes: a DR-Res2net module, which is built based on the Res2Net model. The connection between layers of the Res2Net model is based on the correlation between dense connections and residual connections. The output of the current feature layer is related to the input of the previous feature layer and the feature layer after convolution.

[0013] Assumption: In the DR-Res2net module, the feature layer input to the module is divided into s parts.

[0014] The input from the first feature layer to the i-th feature layer obtained after segmentation is denoted as [x1, x2, ..., x]. i ];

[0015] The intermediate variables from the first feature layer to the i-th feature layer are denoted as [y1, y2, ..., y]. i ];

[0016] The output features from the first feature layer to the ith feature layer are denoted as [z1, z2, ..., z]. i ];

[0017] Where i ≤ s;

[0018] Then, the intermediate variable y i The formula for expressing it is:

[0019]

[0020] in:

[0021] s and i are positive integers, c is convolution; ⊕ is feature addition;

[0022] Output z i The formula for expressing it is:

[0023]

[0024] in:

[0025] s and i are positive integers, and c is the convolution. For concatenation;

[0026] Output features [z1, z2, ..., z i The merged result is output as the final output.

[0027] Its further features are:

[0028] The preprocessing operations in step S2 include: sampling and quantization, pre-emphasis processing, and windowing;

[0029] The sequence of feature vectors to be identified is a 2-dimensional vector. The first dimension is the number of frames after sampling the voiceprint data, and the second dimension is the dimension of the Mel frequency cepstral coefficients.

[0030] The acoustic model is constructed based on ECAPA-TDNN and includes: an input layer, SE-DR-Res2Block, convolutional layer, pooling layer, fully connected layer and AAM-Softmax function connected in sequence.

[0031] After the sequence of feature vectors to be identified is input into the acoustic model, a convolution operation is performed on the input layer, and then it is input into m consecutive SE-DR-Res2Blocks. After being processed by m consecutive N-order SE-DR-Res2Block structures, the processed result is subjected to another convolution operation, followed by a pooling layer. The sequence of feature vectors to be identified is then input into a fully connected layer for classification. Finally, the classification result is processed by AAM-Softmax function vector normalization and then output; where m is a positive integer.

[0032] The SE-DR-Res2Block includes a DR-Res2net module and an SE-Block. A convolutional layer is set before and after the DR-Res2net module. The sequence of feature vectors to be identified input to the SE-DR-Res2Block is processed by the convolutional layer and the DR-Res2net module, and then sent to the SE-Block for feature processing. After adding the input and the processed features, the output is then performed.

[0033] This invention provides a voiceprint recognition method based on the DR-Res2net module. It integrates the characteristics of the dense DenseNet model into the Res2Net model to construct the DR-Res2net model. In the voice recognition model, DR-Res2Block is used as the core module for voiceprint data recognition. During the recognition process, each DR-Res2net module performs residual and dense connections on each output feature simultaneously to obtain richer features. The addition processing in the module increases the amount of information contained in each feature, while the concatenation processing ensures that the features include both high semantic low-resolution and low semantic high-resolution features, maximizing the preservation of features under different receptive fields. This reduces overfitting and improves the classification accuracy of the recognition results. Attached Figure Description

[0034] Figure 1 This is a schematic diagram illustrating the connection relationship of each feature layer in the Res2Block module of the Res2Net model in the prior art.

[0035] Figure 2 This is a schematic diagram of an embodiment of the connection relationship structure of each feature layer in DR-Res2net in this invention;

[0036] Figure 3 This is a schematic diagram of the SE-DR-Res2Block structure;

[0037] Figure 4 This is a schematic diagram of an embodiment of the network structure of the acoustic model in this invention. Detailed Implementation

[0038] This invention provides a voiceprint recognition method based on the DR-Res2net module, which includes the following steps:

[0039] S1: Construct an acoustic model and train it to obtain a trained acoustic model.

[0040] S2: Collect the voiceprint data to be processed, preprocess the original voiceprint signal, and output the voiceprint frame sequence;

[0041] Preprocessing operations include: sampling and quantization, pre-emphasis processing, and windowing.

[0042] S3: Perform time-domain and frequency-domain analysis on the voiceprint frame sequence, extract the Mel frequency cepstral coefficients, and output the feature vector sequence to be identified;

[0043] In this embodiment, the extraction scheme for Mel frequency cepstral coefficients is: 80mfccs

[0044] The sequence of feature vectors to be identified is a 2-dimensional vector. The first dimension is the number of frames after sampling the voiceprint data, and the second dimension is the dimension of the Mel frequency cepstral coefficients.

[0045] S4: After processing the feature vector sequence to be identified, input it into the trained acoustic model for classification and recognition. The recognition result is the classification result of voiceprint recognition.

[0046] The acoustic model in the technical solution of this invention includes: an input layer, an SE-DR-Res2Block layer, a convolutional layer, a pooling layer, a fully connected layer, and an AAM-Softmax function connected in sequence.

[0047] In the technical solution of this invention, the characteristics of the dense DenseNet model are incorporated into the Res2Net model to construct the DR-Res2net model.

[0048] In the DR-Res2net module, the connections between layers are based on correlational connections such as dense connections and residual connections. The output of the current feature layer is related to the input of the previous feature layer and the feature layer after convolution. Therefore, to build a DR-Res2netBlock, the input feature layer needs to be segmented. For example, let's assume that the feature layer input to the DR-Res2net module is segmented into s parts.

[0049] The input from the first feature layer to the i-th feature layer obtained after segmentation is denoted as [x1, x2, ..., x]. i ];

[0050] The intermediate variables from the first feature layer to the i-th feature layer are denoted as [y1, y2, ..., y]. i ];

[0051] The output features from the first feature layer to the i-th feature layer are denoted as [z1, z2, ..., z]. i ];

[0052] Where i ≤ s;

[0053] Then, the intermediate variable y i The formula for expressing it is:

[0054]

[0055] in:

[0056] s and i are positive integers, c is convolution; ⊕ is feature addition (hereinafter abbreviated as add);

[0057] Represents the intermediate output y of the i-th layer i It is the intermediate output y of the previous layer. i-1 After convolution (Conv), batch normalization (BN), and Rectified Linear Unit (ReLU) activation, the signal is compared with the input x of the i-th layer. i The result after adding the feature maps (hereinafter abbreviated as add).

[0058] Output z i The formula for expressing it is:

[0059]

[0060] in:

[0061] s and i are positive integers, and c is the convolution. For superposition The output z of the i-th layer is represented by i y is the intermediate output variable of the i-th layer. i After Conv, BN, and ReLU, and with the intermediate variable y i After adding, then combine with the intermediate variable y i The result is obtained by performing dimension channel concatenation (hereinafter abbreviated as concat) and finally Conv, BN and ReLU processing;

[0062] Output features [z1, z2, ..., z i The merged result is output as the final output.

[0063] In this invention, the dense connection structure of DenseNet is integrated into Res2net as a core module and applied to the acoustic model for speaker recognition. This enables local deepening and expansion of the model. Instead of simply stacking and splicing features, the feature maps of each layer in the channel dimension are connected as input to the next layer, making fuller use of the information from the multi-resolution layers.

[0064] like Figure 2As shown, in each of the DR-Res2net, the input feature sequences are divided into s parts [x1, x2, ..., x] according to the dimension. i After processing, the output features [z1, z2, z3, ..., z] are obtained. i The results are output after merging by dimension. Each feature layer undergoes batch normalization and activation function processing before being processed by the convolutional layers in Equations 1 and 2.

[0065] After the feature vector sequence to be identified is input into the acoustic model, a convolution operation is performed in the input layer, and then it is input into m consecutive SE-DR-Res2Blocks. After being processed by m consecutive N-order SE-DR-Res2Block structures, the processed result is subjected to another convolution operation, followed by a pooling layer. The feature vector sequence to be identified is then input into a fully connected layer for classification. Finally, the classification result is processed by AAM-Softmax function vector normalization and output; where m is a positive integer.

[0066] In the process of audio file feature layer recognition, the original Res2Net model uses residual connections. The technical solution of this invention inserts more hierarchical residual connection structures into the residual units, and combines this with the dense connection characteristics of DenseNet, based on y i z i Expressions; such as Figure 2 In the illustrated embodiment, when S = 4, that is, when i is at most 4:

[0067] The intermediate output variable of the first layer is y. i =x i Final output

[0068] Intermediate output variables of the second layer Final output

[0069] The intermediate output variables of the third layer are Final output

[0070] The fourth layer has no intermediate variables, and the final output depends only on the input feature map. The final output is z4 = x4.

[0071] The DR-Res2Net model structure proposed in this invention, during training, first stacks features from each group through residual connections. Shallow features can directly propagate to deeper features. The identity mapping in this process addresses network degradation to some extent, making information propagation smoother. Then, dense connections are used, allowing each subsequent layer to directly access information from all preceding layers and utilize gradient information, thus enhancing feature transfer and making more efficient use of features. Features are also mutually transferred between groups, preventing the loss of some features.

[0072] Each SE-DR-Res2Block includes: a DR-Res2net module and an SE-Block; a convolutional layer is set before and after the DR-Res2net module; the input feature vector sequence to be identified is processed by the convolutional layer and the DR-Res2net module, then fed into the SE-Block for feature processing, and then the input and the processed features are added together before output.

[0073] SE-Block (Squeeze and Excitation Block) is used after DR-Res2net to reduce the total number of parameters. All module output features are mapped and aggregated, effectively preserving shallow features and capturing more speaker information. This structure deepens the network by stacking more convolutional layers, effectively improving performance, but this significantly increases the model size and computational complexity. Furthermore, the feature information before convolution is not fully utilized, resulting in only limited performance improvement.

[0074] like Figure 3 The diagram shows the connection of each feature layer in SE-DR-Res2Block. The input features are processed using 1x1 Conv, BN, and ReLU methods, and then... Figure 2 The input is processed by a block, followed by 1x1 Conv, BN, and ReLU processing. The processed features are then processed by an SE-Block. Finally, the input and the processed features are added together to obtain the final output.

[0075] In the acoustic model of this invention, the feature vector sequence of the input acoustic model is first processed by a Conv, BN, and ReLU layer in sequence, and then sequentially input into three consecutive SE-DR-Res2Blocks. The size of the convolution kernel in each SE-DR-Res2Block is 3x3, and the expansion rates are 2, 3, and 4, respectively. The result after processing the three SE-DR-Res2Blocks is added, and finally processed by another Conv, BN, and ReLU layer. The final feature is flattened by Attentive Stat Pooling. The processed feature is then passed through a fully connected layer, and finally classified using AAM-Softmax. The classification size is determined according to actual needs.

[0076] The activation function is the Rectified Linear Unit (ReLU) function, and its calculation formula is as follows:

[0077]

[0078] Before audio data is input into the fully connected layer, it needs to be flattened and dimensionality reduced. The flattening and dimensionality reduction process flattens multidimensional data into one-dimensional data in the Attentive Stat Pooling layer.

[0079] Vector normalization is performed using the AAM-Softmax function, the formula for which is:

[0080]

[0081] Where N is the number of training samples, which is taken as 64 in the experiment, θ k,i The angle between the weights of the last fully connected layer and the output of the last fully connected layer (i.e., the learned speaker feature vector). k,i represents the i-th sample in the training samples belonging to the k-th class. k takes values ​​from 1 to the number of classes. In the experiment, the maximum value is 1211 (i.e., k = 1, ..., 1211). m and s are hyperparameters, usually m = 0.2 and s = 30.

[0082] like Figure 4The illustration shows an example of an acoustic model constructed using ECAPA-TDNN as the main body, combined with the DR-Res2net module of this invention. First, the features are passed through a convolutional layer. The result is then input into three consecutive SE-DR-Res2Block structures. Finally, the features processed by the three SE-DR-Res2Blocks are added together and input into a convolutional layer, followed by an Attentive Stat Pooling layer, a fully connected (FC) layer, and finally an AAM-Softmax function. The output of the fully connected layer serves as the speaker vector.

[0083] Based on such Figure 4 The illustrated embodiment uses the Voxceleb1 dataset for experiments. This dataset has 40 categories, and the sample labels are one-hot encoded using PyTorch's `scatter_` function. Based on the actual experimental equipment, S is set to 8, meaning the input acoustic model's feature vector sequence is first processed through a single convolutional layer, then divided into 8 equal parts, processed through several SE-DR-Res2Blocks, and finally convolutional, flattened, and classified.

[0084] The audio frame sequence is analyzed in the time and frequency domains to extract Mel frequency cepstral coefficients and output a feature vector sequence. The sampling frame of the input audio data is 512ms long, with a frame shift of 160ms and 200 sampling frames. The selected scheme for Mel frequency cepstral coefficients is to extract 80-dimensional MFCC features under 80 Mel filter banks. After implementing step S2, the output feature vector sequence is (200, 80).

[0085] The feature vector (200, 80) is input into a 1x1 convolution layer. Then, the feature vector is divided into 8 equal parts (25, 80) and input into three consecutive SE-DR-Res2Block modules. After the same processing is performed three times, it is input into a 1x1 convolution layer again. Finally, convolution, flattening, and classification are performed.

[0086] The features from three SE-DR-Res2Block convolutions are processed by adding, resulting in a feature vector sequence of (1024x3, 80). ReLU activation is performed after each convolution, followed by batch normalization. A statistical pooling layer then flattens the data into one dimension. Finally, a fully connected layer is passed to reduce the parameters of the last fully connected layer, and its output is used as the speaker's feature vector.

[0087] The two-dimensional data, after being processed by three consecutive SE-DR-Res2Block layers, is first flattened into one-dimensional data using a statistical pooling layer, then enters the first fully connected layer with 192 neurons; finally, it enters the second fully connected layer with 40 neurons, meaning this classification is for 40 categories. Finally, it is processed by the normalized exponential function AAM-Softmax to output the final classification result.

[0088] The experiments were conducted in a Windows 10 system environment with a GTX 2060 graphics card, an i7-9700 CPU, and 16GB of RAM. PyTorch was used as the deep learning framework, and the Voxceleb1 standard dataset for voiceprint recognition was employed for performance testing. Equal error rate (EER) and minimum detection cost functions (DCF0.1, DCF0.01, DCF0.001) were used as performance metrics to evaluate performance. x-vector was used as the baseline system.

[0089] The standard dataset Voxceleb1 contains 1211 items in the train set, and each item is divided into a training set and a test set in a 7:3 ratio. The test set consists of 40 WAV audio files in the test set, which are divided into a registration set and a validation set according to the files provided by Voxceleb1.

[0090] First, experiments on speaker recognition were conducted on the Res2Net-50 network using different Res2NetBlock modules. The detection systems used in the following experiments were: the original Res2Net-50 system, a system that applies the FULL-Res2Net structure proposed by Chen Zhigao et al. to Res2Net-50 (hereinafter referred to as FULL-Res2Net-50), the DR-Res2Net-50 system of the present invention, and the FULL-DR-Res2Net-50 system, a variant of FULL-Res2Net. The dataset selected was voxceleb1, and the specific results are shown in Table 1 below:

[0091] Table 1 Performance comparison of Voxceleb1 test set on different Res2Net-50 systems

[0092]

[0093] As shown in Table 1, the DR-Res2Net-50 system shows a 5.9% decrease in EER compared to the Res2Net-50 system, and the FULL-DR-Res2Net-50 system shows a 5.4% decrease in EER compared to the FULL-Res2Net-50 system. The minimum detection cost function is also reduced, effectively demonstrating the effectiveness of combining dense connections and residual connections. The system containing the module proposed in this paper has the best performance.

[0094] By using different Res2NetBlocks in the main network of ECAPA-TDNN, the detection systems used in the following experiments are as follows: the original Res2Net structure applied to the ECAPA-TDNN system, abbreviated as Res2Net in the table; the FULL-Res2Net structure proposed by Chen Zhigao et al. applied to ECAPA-TDNN, abbreviated as FULL-Res2Net; the DR-Res2Net system of the technical solution of this invention; and the FULL-DR-Res2Net system, a variant of FULL-Res2Net, for voiceprint recognition classification experiments. The specific results are shown in Table 2 below.

[0095] Table 2 Performance comparison of Voxceleb1 test set on different ECAPA-TDNN systems

[0096]

[0097] As shown in Table 2, the DR-Res2Net-based architecture, compared to the original ECAPA-TDNN system, achieves a 10% relative decrease in EER with only a 13% increase in parameters. The minDCF (0.1 and 0.01) also shows a 9% relative decrease. For both FULL-Res2Net and FULL-DR-Res2Net systems, incorporating the proposed architecture results in a 5.5% relative decrease in EER and a reduction in minDCF. The experimental results demonstrate the effectiveness of combining dense and residual connections, with the proposed DR-Res2Net architecture achieving the best results.

[0098] Table 3 Performance comparison of the SITW test set under different ECAPA-TDNN systems

[0099]

[0100] To verify the effectiveness of the architecture, this experiment was conducted in the core-core test scenario of the SITW dataset. The architecture was trained using the voxceleb1 training set and tested using the SITW test set. The experimental results are shown in Table 3. DR-Res2Net shows a 6.6% decrease in EER compared to the original architecture. FULL-DR-Res2Net shows a 4.3% decrease in EER compared to the FULL-Res2Net architecture.

[0101] To evaluate the system's performance on voiceprint data of different durations, this experiment used three sub-test sets under the core-core test set: speech less than 15 seconds, speech greater than 15 seconds but less than 25 seconds, and speech greater than 25 seconds but less than 40 seconds.

[0102] Table 4. EER (%) under different SITW durations

[0103] system <15s 15-25s 25-40s x-vector 7.52 7.21 6.65 Res2Net 4.56 4.21 3.43 FULL-Res2Net 4.75 4.30 3.52 DR-Res2Net 4.06 3.82 3.30 FULL-DR-Res2Net 4.12 4.07 3.42

[0104] The experimental results are shown in Table 4. DR-Res2Net exhibits the best performance across all speech durations. Furthermore, the EER performance of all systems improves with increasing duration, indicating that longer speech durations contain more speaker information. Systems employing DR-Res2Net show improved performance compared to the original systems, with the performance improvement being more significant for shorter durations. DR-Res2Net demonstrates the best performance across all durations. This result indicates that the technical solution of this invention is effective for different speech durations, and is particularly effective for short-duration speech.

[0105] In summary, the technical solution provided by this invention can provide richer and more effective feature information when processing voiceprint data, ensuring that the model has stronger generalization ability, good equal error rate and minimum detection cost function.

[0106] This invention provides a voiceprint recognition method based on the DR-Res2net module, applying the DR-Res2Net structure, a variant of the Res2Net structure, to the field of voiceprint recognition. Compared to the traditional Res2Net structure, it integrates the advantages of DenseNet, not simply stacking and splicing features, but aggregating features from each layer while retaining shallow feature information. This allows features at different scales and receptive fields to complement each other, maximizing the utilization of features from different layers.

Claims

1. A voiceprint recognition method based on the DR-Res2net module, comprising the following steps: S1: Construct an acoustic model and train the acoustic model to obtain a trained acoustic model; S2: Collect the voiceprint data to be processed, preprocess the original voiceprint signal, and output the voiceprint frame sequence; S3: Perform time-domain and frequency-domain analysis on the voiceprint frame sequence, extract the Mel frequency cepstral coefficients, and output the feature vector sequence to be identified; S4: After processing, the sequence of feature vectors to be identified is input into the trained acoustic model for classification and recognition. The resulting recognition result is the classification result of voiceprint recognition. Its features are: The acoustic model includes: a DR-Res2net module, which is built based on the Res2Net model. The connection between layers of the Res2Net model is based on the correlation between dense connections and residual connections. The output of the current feature layer is related to the input of the previous feature layer and the feature layer after convolution. In the DR-Res2net module, the feature layer input to the module is divided into s parts. The input from the first feature layer to the i-th feature layer obtained after segmentation is denoted as [x1, x2, ..., x]. i ]; The intermediate variables from the first feature layer to the i-th feature layer are denoted as [y1, y2, ..., y]. i ]; The output features from the first feature layer to the ith feature layer are denoted as [z1, z2, ..., z]. i ]; Where i ≤ s; Then, the intermediate variable y i The formula for expressing it is: in: s and i are positive integers, c is convolution; ⊕ is feature addition; Output z i The formula for expressing it is: in: s and i are positive integers, and c is the convolution. For concatenation; Output features [z1, z2, ..., z i The merged result will be output as the final output. Represents the intermediate output y of the i-th layer i It is the intermediate output y of the previous layer. i-1 After convolution processing, batch normalization, and Rectified Linear Unit activation, the input x of the i-th layer is compared with the input x. i The result after adding the feature maps using Addition; The output z of the i-th layer is represented by i y is the intermediate output variable of the i-th layer. i After Conv, BN, and ReLU, and with the intermediate variable y i After adding, then combine with the intermediate variable y i The result is obtained by performing dimension channel concatenation processing, followed by Conv, BN, and ReLU processing. During training, the DR-Res2Net model first stacks features through residual connections, allowing shallow features to propagate directly to deep features. The identity mapping in this process makes the information propagation smoother. Then, dense connections are formed, so that each subsequent layer can directly obtain information from all previous layers, and each layer can directly utilize gradient information, which enhances the transmission of features.

2. The voiceprint recognition method based on the DR-Res2net module according to claim 1, characterized in that: The preprocessing operations in step S2 include: sampling and quantization, pre-emphasis processing, and windowing.

3. The voiceprint recognition method based on the DR-Res2net module according to claim 1, characterized in that: The sequence of feature vectors to be identified is a 2-dimensional vector. The first dimension is the number of frames after sampling the voiceprint data, and the second dimension is the dimension of the Mel frequency cepstral coefficients.

4. The voiceprint recognition method based on the DR-Res2net module according to claim 1, characterized in that: In the acoustic model, the DR-Res2net module has a convolutional layer before and after it, and then an SE-Block is set at the end to form an SE-DR-Res2Block. The input feature vector sequence to be identified in the SE-DR-Res2Block is processed by the convolutional layer and the DR-Res2net module, then fed into the SE-Block for feature processing. After adding the input and the processed features, the output is generated.

5. The voiceprint recognition method based on the DR-Res2net module according to claim 4, characterized in that: The acoustic model is constructed based on ECAPA-TDNN and includes: an input layer, SE-DR-Res2Block, convolutional layer, pooling layer, fully connected layer and AAM-Softmax function connected in sequence. After the sequence of feature vectors to be identified is input into the acoustic model, a convolution operation is performed on the input layer, and then it is input into m consecutive SE-DR-Res2Blocks. After being processed by m consecutive N-order SE-DR-Res2Block structures, the processed result is subjected to another convolution operation, followed by a pooling layer. The sequence of feature vectors to be identified is then input into a fully connected layer for classification. Finally, the classification result is processed by AAM-Softmax function vector normalization and then output; where m is a positive integer.

Images