A multi-state voiceprint feature recognition method and device

Abstract

The present invention relates to a multi-state voiceprint feature recognition method and device. The method comprises: obtaining an audio signal, preprocessing the audio signal, obtaining a speech feature set, inputting the speech feature set into a voiceprint recognition network model, obtaining multiple voiceprint feature states, calculating similarity of the voiceprint feature states according to the multiple voiceprint feature states of the speech feature set, updating the current voiceprint feature state, iteratively updating the parameters of the voiceprint recognition network model according to the voiceprint feature state, until the voiceprint recognition network model converges. The present invention calculates the similarity of multi-state voiceprint features of audio signals, updates network parameters through loss functions, and ensures the continuity of voiceprint feature state transitions during the network parameter update process, thereby associating different states of similar or identical audio signals and accurately identifying them.

Description

Technical Field

[0001] The present invention relates to the field of speech recognition, and in particular to a multi-state voiceprint feature recognition method and device. Background Art

[0002] With the rapid development of machine learning, voiceprint recognition technology has great potential and urgent needs in financial anti-fraud and other areas. Deep learning adopts a multi-layer neural layer structure to extract information from the raw data of the input layer by layer, automatically mines complex complex representation features in large data sets, and avoids artificial feature engineering. Deep belief networks, stacked denoising autoencoders, and long short-term memory networks have been applied, and have good dynamic learning effects compared to traditional machine learning methods. Convolutional neural networks, as a distributed parallel deep network model, have also been widely used in fault diagnosis.

[0003] Compared with other deep learning methods, the time-delay neural network (TNDD) was first proposed by Hinton in 1989 to solve the problem of phoneme recognition. However, this structure only considers local contextual information, while ECAPA-TDNN considers global information, so it is more suitable for extracting voiceprint features over a period of time. However, the voiceprint features of the same person in different time periods or under different physical health conditions will be somewhat different. Therefore, it is necessary to process and analyze the correlation of voiceprint features in this case. In addition, after the existing voiceprint model is trained, the voiceprint differentiation between people with the same pronunciation method needs to be improved, which limits the promotion of voiceprint recognition technology. Summary of the invention

[0004] In order to solve the problem that the above-mentioned prior art cannot recognize the same voice with high accuracy, the present invention provides a multi-state voiceprint feature recognition method and device. The technical solution is as follows: A multi-state voiceprint feature recognition method includes the following steps:

[0005] S1, acquiring an audio signal, and preprocessing the audio signal to obtain a speech feature set;

[0006] S2, inputting the speech feature set into the voiceprint recognition network model to obtain a plurality of voiceprint feature states, wherein the voiceprint feature states include a current voiceprint feature state and a historical voiceprint feature state;

[0007] S3, calculating similarity of the voiceprint feature states according to the multiple voiceprint feature states of the audio signal in the speech feature set, and updating the current voiceprint feature state,

[0008] S4, iteratively updating the parameters of the voiceprint recognition network model according to the number of the voiceprint feature states until the voiceprint recognition network model converges.

[0009] Based on the above method, the present invention can also make the following improvements.

[0010] Furthermore, the audio signal is preprocessed in S1 as follows: after acquiring the audio signal, the audio signal is sequentially subjected to pre-emphasis processing, frame processing, windowing processing, fast Fourier transform processing, triangular bandpass filter, logarithmic operation and dynamic differential parameter operation to obtain the speech feature set.

[0011] Furthermore, S1 also includes: selecting a fixed length of speech features from the speech feature set and inputting the fixed length into the voiceprint recognition network model, where the fixed length is T.

[0012] Furthermore, before S1, the method further includes: performing noise reduction and dereverberation processing on the audio signal.

[0013] Further, the similarity calculation of the voiceprint feature state and updating the current voiceprint feature state in S3 is specifically as follows: the similarity calculation of the current voiceprint feature state and the historical voiceprint feature state is performed;

[0014] If the number of current voiceprint feature states is less than N and the distance between the current voiceprint feature state and the historical voiceprint feature state is greater than a first threshold, then a new voiceprint feature state is added;

[0015] If the number of current voiceprint feature states is less than N and the distance between the current voiceprint feature state and the historical voiceprint feature state is less than a first threshold, the historical voiceprint feature state and the current voiceprint feature state are merged and counted;

[0016] If the number of current voiceprint feature states is equal to N and the distance between the current voiceprint feature state and the historical voiceprint feature state is greater than a first threshold, an inter-class distance calculation is performed on the outlier voiceprint feature state or the newly added voiceprint feature state.

[0017] If the outlier voiceprint feature state or the newly added voiceprint feature state has a high similarity with the historical voiceprint feature state, then deleting the outlier voiceprint feature state or the newly added voiceprint feature state;

[0018] If the outlier voiceprint feature state or the newly added voiceprint feature state has a low similarity to the historical voiceprint feature state, the distances between the N voiceprint feature states are updated and the two voiceprint feature states with the smallest distances between them are merged.

[0019] Further, the S4 is specifically: using the speech feature set to iteratively train the voiceprint recognition network model for multiple times, performing AAM-Softmax calculation according to the number of voiceprint feature states or the mean of the voiceprint feature states, and when the standard deviation abs=(r2-x2) is less than the second threshold, it is determined that the initially trained voiceprint recognition network model has converged, where r is the classification probability of the voiceprint feature state, and x is the initial expected target parameter.

[0020] Furthermore, the voiceprint recognition network model includes: an input layer, a feature extraction layer and an output layer, wherein the input layer is a time-delay neural network structure, which is used to perform pre-input processing on the speech feature set;

[0021] The feature extraction layer includes an SE residual network layer and a feature fusion layer. After the speech feature set is pre-input processed by the input layer, the input speech feature is processed by the SE residual network layer for feature dimension reduction and dimension recovery to obtain a speech feature vector. The speech feature vector is subjected to multi-layer feature calculation by the feature fusion layer to obtain a voiceprint feature vector.

[0022] The output layer includes an attention statistical pooling layer and an AAM-Softmax layer. The voiceprint feature vector is pooled for voiceprint classification probability through the attention statistical pooling layer. The AAM-Softmax layer is used to classify the audio signal according to the voiceprint classification probability.

[0023] Another object of the present invention is to provide a multi-state voiceprint feature recognition device, comprising a memory and a processor, wherein the memory stores at least one program, and the at least one program is executed by the processor to implement the above-mentioned multi-state voiceprint feature recognition method.

[0024] Another object of the present invention is to provide a computer-readable storage medium, including a memory, wherein the memory stores a computer program, and when the computer program is executed by a processor, the multi-state voiceprint feature recognition method as described above is implemented.

[0025] The beneficial effect of the present invention is that by calculating the similarity of multi-state voiceprint features of audio signals and updating the network parameters through the loss function, the continuity of the voiceprint feature state transition during the network parameter update process is ensured, thereby associating different states of similar or identical audio signals (audio signals emitted by the same object or the same person) and accurately identifying them. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A flow chart of a multi-state voiceprint feature recognition method of the present invention;

[0027] Figure 2 It is a flow chart of the feature extraction method of audio signal;

[0028] Figure 3 This is the flow chart of the network back propagation algorithm;

[0029] Figure 4 Schematic diagram of the voiceprint recognition network model structure. DETAILED DESCRIPTION

[0030] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0031] like Figure 1 As shown, the present invention provides a multi-state voiceprint feature recognition method and device, comprising the following steps:

[0032] S1, acquiring an audio signal, and preprocessing the audio signal to obtain a speech feature set;

[0033] S2, inputting the speech feature set into the voiceprint recognition network model to obtain a plurality of voiceprint feature states, wherein the voiceprint feature states include a current voiceprint feature state and a historical voiceprint feature state;

[0034] S3, calculating similarity of the voiceprint feature states according to the multiple voiceprint feature states of the audio signal in the speech feature set, and updating the current voiceprint feature state;

[0035] S4, iteratively updating the parameters of the voiceprint recognition network model according to the number of the voiceprint feature states until the voiceprint recognition network model converges.

[0036] The present invention calculates the similarity of multi-state voiceprint features of audio signals, updates network parameters through a loss function, and ensures the continuity of voiceprint feature state transitions during the network parameter update process, thereby associating different states of similar or identical audio signals (audio signals emitted by the same object or the same person) and accurately identifying them.

[0037] Specifically, in this embodiment, the audio signal is extracted into speech features such as Figure 2 As shown, by performing noise reduction and dereverberation processing on the audio signal, the instability of the audio signal caused by force majeure factors can be avoided as much as possible, and the quality of the audio signal can be improved. The speech features in the speech feature set are selected with a fixed length and input into the voiceprint recognition network model, and the fixed length is T. In the present invention, the MFCC parameters in the speech feature parameters are used to extract data from the audio signal to obtain the speech feature set.

[0038] After acquiring the audio signal, the audio signal is extracted into speech features such as Figure 2 As shown, the audio signal is subjected to pre-emphasis processing, frame processing, windowing processing, fast Fourier transform processing, triangular bandpass filter, logarithmic operation and dynamic differential parameter operation in sequence to obtain the speech feature sequence set. Pre-emphasis processing of the audio signal is actually to pass the audio signal through a high-pass filter:

[0039] H(Z)=1-μz -1 ;

[0040] Where H(Z) is the audio signal, and the value of μ is between 0.9 and 1.0. By pre-emphasizing the high-frequency part of the audio signal, the spectrum of the signal can be flattened and kept in the entire frequency band from low frequency to high frequency, and the spectrum can be obtained with the same signal-to-noise ratio. At the same time, it is also to eliminate the effects of the vocal cords and lips during the occurrence process, to compensate for the high-frequency part of the audio signal suppressed by the pronunciation system, and to highlight the high-frequency resonance peak. The audio signal after pre-emphasis is framed, and N (N is a natural number) sampling points are preset to form a frame. It can avoid an overlapping area between adjacent frames due to excessive changes in two adjacent frames. Then, each frame in the audio signal is windowed, that is, each frame in the audio signal is multiplied by a Hamming window to increase the continuity of the left and right ends of the frame, so that the audio signal roughly presents a periodic change.

[0041] Since it is usually difficult to see the characteristics of the signal from the transformation in the time domain, it is usually converted into energy distribution in the frequency domain for observation. Different energy distributions can represent the characteristics of different voices. Therefore, after each frame in the audio signal is multiplied by the Hamming window, each frame in the audio signal must also undergo a fast Fourier transform to obtain the energy distribution in the spectrum. The frequency spectrum of each frame is obtained by performing a fast Fourier transform on the framed and windowed signals. The power spectrum of the audio signal is obtained by taking the modulus square of the frequency spectrum of the audio signal. The power spectrum of the audio signal is smoothed by passing it through a triangular filter bank, while eliminating the harmonics in the power spectrum, highlighting the resonance peak of the original voice and reducing the amount of calculation.

[0042] In addition, the logarithmic energy of the important feature of speech is calculated by the volume (i.e. energy) of each frame in the audio signal. The dynamic characteristics of speech in the audio signal are described by extracting dynamic differential parameters. The speech characteristics of each frame are processed in multiple dimensions, and the dynamic and static characteristics are combined, which can effectively improve the recognition performance of the recognition network when the speech feature sequence set is used for recognition training.

[0043] Specifically in this embodiment, the similarity between the current voiceprint feature state and the historical voiceprint feature state is calculated, for example, the cosine similarity is calculated; if the number of the current voiceprint feature state is less than N and the distance between the current voiceprint feature state and the historical voiceprint feature state is greater than a first threshold, a new voiceprint feature state is added; if the number of the current voiceprint feature state is less than N and the distance between the current voiceprint feature state and the historical voiceprint feature state is less than the first threshold, the historical voiceprint feature state and the current voiceprint feature state are merged and counted;

[0044] If the number of current voiceprint feature states is equal to N and the distance between the current voiceprint feature state and the historical voiceprint feature state is greater than a first threshold, the inter-class distance calculation is performed on the outlier voiceprint feature state or the newly added voiceprint feature state. If the outlier voiceprint feature state or the newly added voiceprint feature state has a high similarity with the historical voiceprint feature state, the outlier voiceprint feature state or the newly added voiceprint feature state is deleted; if the outlier voiceprint feature state or the newly added voiceprint feature state has a low similarity with the historical voiceprint feature state, the distance between the N voiceprint feature states is updated and the two voiceprint feature states with the smallest distance between the voiceprint feature states are merged.

[0045] Specifically in this embodiment, the voiceprint recognition network model is iteratively trained multiple times using the speech feature set, and AAM-Softmax calculation is performed according to the number of voiceprint feature states or the mean of voiceprint feature states. When the standard deviation abs=(r2-x2) is less than a second threshold, it is determined that the initially trained voiceprint recognition network model has converged, where r is the classification probability of the voiceprint feature state, and x is the initial expected target parameter.

[0046] The goal of evaluating and optimizing the effect of the network model is to make a prediction of the label space of the sample mapped to the fully connected layer network features, and then compare it with the true value to measure the distance between the predicted value and the true sample label, and guide the network parameter learning and representation learning through error back propagation. For multi-classification tasks, the cross-entropy loss function is generally used, which is defined as the cross entropy between the true value and the predicted value. The cross entropy represents the distance between the actual output and the expected output. The smaller the cross entropy value, the closer the two are. For a given training set Input all samples into the network and get the network output as Then the objective function on data set A is:

[0047]

[0048] Where y n is the vector representation of the label corresponding to the nth data. The learning parameters are iteratively updated and optimized by minimizing the objective function. The smaller the Loss, With yn The closer the predicted value is, the closer it is to the true value.

[0049] like Figure 3 As shown in the figure, the back propagation algorithm implements an iterative process. At the beginning of each iteration, a small batch of training data needs to be selected first. This small batch of data is called a batch. Then, this batch will use the weights and thresholds initialized by the current network, from the input layer to the hidden layer, and finally reach the output layer to obtain the prediction result. Since the training data and the real annotations correspond one-to-one, the distance between the predicted value and the true value of the current model can be calculated, and the predicted result and the true value can be used to form a cost function. This process is the forward propagation process. Through the chain rule, the derivative of the loss function with respect to each parameter can be calculated, and the partial derivative of the objective function with respect to the net input of the neurons in the first layer is defined as the error term as an intermediate variable. Since the error term of a neuron in the first layer is the weighted sum of the error terms of all neurons in the 1+1 layer connected to it, and then multiplied by the gradient of the activation function of the neuron. The error can be back-propagated from the output layer to the hidden layer until it propagates to the input layer. In the process of back propagation, the values ​​of each parameter are adjusted according to the error, and the above process is continuously iterated until the target converges and the loss function reaches the optimal value. Based on the gap between the predicted value and the true value, the back propagation algorithm will update the model parameters accordingly, making the predicted result of the model in this batch closer to the true label. Once the network structure of the objective function is determined, automatic differentiation can be used to calculate the gradient of the objective function with respect to each parameter, and then an optimization algorithm is selected. Using the gradient-based optimization algorithm, the optimal value of the objective function is obtained and the optimization parameters are iteratively updated.

[0050] Specifically in this embodiment, Figure 4The input layer, feature extraction layer and output layer shown in the figure, the input layer is a time-delay neural network structure, that is, a one-dimensional neural network convolution layer, which is used to perform pre-input processing on the speech feature set, including an activation function, and the activation function is mainly used to perform nonlinear mapping operations on the output results of the convolution layer. Currently, there are mainly functions such as ReLU, sigmoid, and tanh. The present invention uses the ReLU (Rectified Linear Unit) rectified linear unit as a preferred embodiment. Within the specific range of the deep neural network, it has the characteristics of nonlinearity and linearity. For all inputs greater than 0, the corresponding gradient has a constant derivative value, which greatly enhances the information integration capability and speeds up and simplifies the training of the network. In addition, ReLU can effectively reduce the gradient vanishing and overfitting problems of the CNN model; the feature extraction layer includes an SE residual network layer and a feature fusion layer, wherein the SE residual network layer includes multiple SE residual network structures, and the SE residual network structure mainly includes two parts: Squeeze (compression) and Excitation (excitation), wherein the speech feature set is pre-input processed by the input layer, and the input speech features are subjected to feature dimension reduction using the Squeeze (compression) part in the SE residual network layer, specifically, for example, W, H represent feature width and height. C represents the number of channels, the input feature size is W×H×C, and the global average pooling of the compression part in the SE residual network layer, after the compression operation, the feature is compressed into a 1×1×C vector. The compressed feature vector is then passed through a Sigmoid (activation function) gate to obtain a normalized weight between 0 and 1. Finally, a Scale operation is performed to weight the normalized weight to the features of each channel to perform dimensionality recovery to obtain a speech feature vector. This can make the features more nonlinear and better fit the complex correlations between channels. At the same time, the number of parameters and the amount of calculation are greatly reduced.

[0051] The SE residual network layer in the present invention includes multiple SE residual network structures. The following adopts three layers of SE-Res2Block. The first layer of SE-Res2Block outputs a speech feature vector C1, the second layer of SE-Res2Block network structure outputs a speech feature vector C2, and the third layer of SE-Res2Block outputs a speech feature vector C3. The speech feature vectors C1, C2, and C3 are subjected to multi-layer feature calculation in the feature fusion layer to obtain a voiceprint feature vector. The voiceprint feature vector is pooled for voiceprint classification probability through the attention statistical pooling layer. The output layer includes an output layer and a classification layer. The AAM-Softmax layer is used to classify the audio signal according to the voiceprint classification probability, that is, it is used for classification problems. Its expression is to replace the activation function with the AAM-Softmax function on the basis of the fully connected layer. The activity value of the previous neural network layer is used as the input layer of the AAM-Softmax layer. The output of the AAM-Softmax layer has K neurons corresponding to K types of sample labels. The class scores are calculated through the AAM-Softmax layer. The normalized exponential function is used to represent the probability distribution of different categories of the current batch of samples. The output with the largest classification probability is taken as the final prediction.

[0052] The data processing in the time-delay neural network structure, attention statistics pooling layer and output layer also includes batch normalization (BN). BN is an effective layer-by-layer normalization method. During the forward propagation of batch data in the model, the parameter distribution of each layer of the model changes, which leads to a decrease in the training speed. At the same time, it requires good parameter initialization, which makes it very difficult to train models with saturated nonlinearity. Therefore, it is necessary to perform batch normalization on the hidden layer in the neural network. During each training of the model, the input batch data is scaled and translated to a stable mean and standard deviation through transformation, so that each dimension of the net input is normalized to a standard normal distribution, thereby improving the optimization efficiency without paying special attention to the initialization of parameters. Although batch normalization can also be applied to the activity value of neurons, batch normalization of net input is more conducive to network optimization. Therefore, in practical applications, BN is generally used after affine transformation and before activation function. Applying the BN layer can enhance the model's adaptability to different input distributions and has a slight regularization effect. It not only speeds up the model's convergence and training speeds, but also effectively improves the model's convergence rate. It can also alleviate the "gradient diffusion" effect during deep network training to a certain extent, and improve the generalization performance of the network, making it easier and more stable to train deep network models. In addition, BN can also be used to speed up the model's training speed and even improve the model's accuracy.

[0053] Another object of the present invention is to provide a multi-state voiceprint feature recognition device, comprising a memory and a processor, wherein the memory stores at least one program, and the at least one program is executed by the processor to implement the above-mentioned multi-state voiceprint feature recognition method.

[0054] Another object of the present invention is to provide a computer-readable storage medium, including a memory, wherein the memory stores a computer program, and when the computer program is executed by a processor, the multi-state voiceprint feature recognition method as described above is implemented.

[0055] The reader should understand that in the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples without contradiction.

[0056] 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 method embodiments described above are only illustrative, for example, the division of steps is only a logical function division, and there may be other division methods in actual implementation, such as multiple steps can be combined or integrated into another step, or some features can be ignored or not executed.

[0057] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A multi-state voiceprint feature recognition method, characterized in that: The following steps are included: S1, obtaining an audio signal, and preprocessing the audio signal to obtain a speech feature set; S2, inputting the speech feature set into the voiceprint recognition network model to obtain a plurality of voiceprint feature states, wherein the voiceprint feature states include a current voiceprint feature state and a historical voiceprint feature state; S3, calculating similarity of the voiceprint feature states according to the multiple voiceprint feature states of the speech feature set, and updating the current voiceprint feature state; S4, iteratively updating the parameters of the voiceprint recognition network model according to the voiceprint feature state until the voiceprint recognition network model converges; The similarity calculation of the voiceprint feature state and updating the current voiceprint feature state in step S3 are specifically as follows: Calculate the similarity between the current voiceprint feature state and the historical voiceprint feature state; If the number of current voiceprint feature states is less than N and the distance between the current voiceprint feature state and the historical voiceprint feature state is greater than a first threshold, then a new voiceprint feature state is added; If the number of current voiceprint feature states is less than N and the distance between the current voiceprint feature state and the historical voiceprint feature state is less than a first threshold, the historical voiceprint feature state and the current voiceprint feature state are merged and counted; If the number of current voiceprint feature states is equal to N and the distance between the current voiceprint feature state and the historical voiceprint feature state is greater than a first threshold, then the inter-class distance calculation is performed on the outlier voiceprint feature state or the newly added voiceprint feature state. If the outlier voiceprint feature state or the newly added voiceprint feature state has a high similarity with the historical voiceprint feature state, then deleting the outlier voiceprint feature state or the newly added voiceprint feature state; If the outlier voiceprint feature state or the newly added voiceprint feature state has a low similarity to the historical voiceprint feature state, the distances between the N voiceprint feature states are updated and the two voiceprint feature states with the smallest distances between them are merged.

2. The multi-state voiceprint feature recognition method according to claim 1, characterized in that: The preprocessing of the audio signal in S1 is specifically as follows: after acquiring the audio signal, the audio signal is sequentially subjected to pre-emphasis processing, frame processing, windowing processing, fast Fourier transform processing, triangular bandpass filter, logarithmic operation and dynamic differential parameter operation to obtain the speech feature set.

3. The multi-state voiceprint feature recognition method according to claim 1, characterized in that: The S1 also includes: selecting a fixed length of speech features from the speech feature set and inputting the selected speech features into the voiceprint recognition network model.

4. The multi-state voiceprint feature recognition method according to claim 1, characterized in that: Before S1, the step further includes: performing noise reduction and dereverberation processing on the audio signal.

5. The multi-state voiceprint feature recognition method according to claim 1, characterized in that: The S4 is specifically: using the speech feature set to iteratively train the voiceprint recognition network model for multiple times, performing AAM-Softmax calculation according to the number of voiceprint feature states or the mean of the voiceprint feature states, and when the standard deviation abs=(r2–x2) is less than the second threshold, it is determined that the initially trained voiceprint recognition network model has converged, where r is the classification probability of the voiceprint feature state, and x is the initial expected target parameter.

6. The multi-state voiceprint feature recognition method according to claim 2, characterized in that: The voiceprint recognition network model includes: an input layer, a feature extraction layer and an output layer, wherein the input layer is a time-delay neural network structure, which is used to perform pre-input processing on the speech feature set; The feature extraction layer includes an SE residual network layer and a feature fusion layer. After the speech feature set is pre-input processed by the input layer, the input speech feature is processed by the SE residual network layer for feature dimension reduction and dimension recovery to obtain a speech feature vector. The speech feature vector is subjected to multi-layer feature calculation by the feature fusion layer to obtain a voiceprint feature vector. The output layer includes an attention statistical pooling layer and an AAM-Softmax layer. The voiceprint feature vector is pooled for voiceprint classification probability through the attention statistical pooling layer. The AAM-Softmax layer is used to classify the audio signal according to the voiceprint classification probability.

7. A multi-state voiceprint feature recognition device, characterized in that: The method comprises a memory and a processor, wherein the memory stores at least one program, and the at least one program is executed by the processor to implement the multi-state voiceprint feature recognition method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that: The method comprises a memory, wherein a computer program is stored in the memory, and when the computer program is executed by a processor, the multi-state voiceprint feature recognition method according to any one of claims 1 to 6 is implemented.

Images