Voiceprint recognition method and system based on wavelet transform convolution, terminal and medium

Abstract

The invention provides a voiceprint recognition method and system based on wavelet transform convolution, a terminal and a medium. The method comprises the steps of obtaining a to-be-recognized voice of a current speaker; extracting Mel-frequency cepstral coefficient features based on the voiceprint features of the voice to be recognized; inputting the Mel-frequency cepstral coefficient features into a trained voiceprint recognition model so as to perform feature extraction on the Mel-frequency cepstral coefficient features, and obtaining voiceprint feature embedding vectors; and identifying the voiceprint feature embedded vector to judge whether the current speaker is a registered user or not. According to the method, the multi-dimensional hybrid wavelet transform convolution, the attention mechanism and the Res2Net network structure are organically combined, the capturing capability of different frequency features is remarkably improved, a multi-segment voice similarity calculation strategy is further adopted, the adaptability to complex environments and noise is improved, and the recognition precision of voiceprint features is effectively enhanced.

Description

Technical Field

[0001] The present application relates to the field of speech recognition technology, and in particular to a voiceprint recognition method, system, terminal and medium based on wavelet transform convolution. Background Art

[0002] Speech recognition, a technology that converts human speech into text or commands, is widely used in various terminal devices. Voiceprint recognition, among other things, determines the speaker's identity by analyzing their voice signals. It relies on unique information about each person's vocal tract, such as resonant frequency, timbre, pitch, and pronunciation habits. These characteristics are influenced by physiological structure and speaking style, making them unique and stable. The process involves voiceprint collection, feature extraction, model training, and matching evaluation. It is widely used in finance, security, and other fields, such as telephone banking authentication, smart home control, and security monitoring, providing a convenient and secure solution for identity verification.

[0003] Currently, existing voiceprint recognition technology solutions typically employ a residual network architecture to construct a speaker embedding layer model. During the residual calculation process, convolution and deconvolution operations are performed to ensure that the output features of the residual module are consistent with the input features in terms of time scale, thus ensuring that important information is not lost in the time domain. However, existing models, using a simple convolution operator, cannot effectively distinguish and highlight key frequency components relevant to the task and are susceptible to interference from noise and redundant frequencies, resulting in inaccurate feature extraction. Furthermore, standard convolution cannot perform multi-scale feature decomposition and lacks the ability to differentiate the importance of different time and frequency dimensions in the feature graph. This makes it unable to effectively handle the complex variations of speech features in the time and frequency domains, resulting in limited speaker recognition accuracy. Furthermore, existing models have a large number of parameters and computational requirements, making them unsuitable for deployment on embedded devices, such as those with resource-constrained hardware environments, thus limiting their scope of application. Therefore, accurately identifying voiceprints is a technical problem that those skilled in the art urgently need to address. Summary of the Invention

[0004] The present application provides a voiceprint recognition method, system, terminal and medium based on wavelet transform convolution, which are used to solve the technical problem of how to accurately identify voiceprints.

[0005] A first aspect of the present application provides a voiceprint recognition method based on wavelet transform convolution, the method comprising: obtaining a speech to be recognized of a current speaker;

[0006] Extracting Mel-frequency cepstral coefficient features based on the voiceprint characteristics of the speech to be recognized;

[0007] Inputting the Mel-cepstral coefficient features into a trained voiceprint recognition model to extract features from the Mel-cepstral coefficient features and obtain a voiceprint feature embedding vector, including: performing multi-dimensional mixed wavelet transform convolution processing and frequency weight calibration processing on the Mel-cepstral coefficient features to extract time-frequency features; converting frame-level features in the time-frequency features into sentence-level features to generate a voiceprint feature embedding vector; and

[0008] The voiceprint feature embedding vector is identified to determine whether the current speaker is a registered user.

[0009] In some embodiments of the first aspect of the present application, extracting the time-frequency features includes:

[0010] Performing a two-dimensional wavelet transform convolution process on the Mel-cepstral coefficient feature based on a two-dimensional wavelet transform convolution layer to perform multi-scale time-frequency decomposition on the Mel-cepstral coefficient feature to obtain multi-scale local time-frequency features;

[0011] Grouping the multiple feature channels of the multi-scale local time-frequency features based on a two-dimensional residual network, and performing convolution processing on the multi-scale local time-frequency features of each group of feature channels to obtain multi-channel multi-scale local time-frequency features;

[0012] Based on the frequency attention network, a weight calibration process is performed on the frequency dimension of the multi-channel and multi-scale local time-frequency features to obtain multi-channel and multi-scale local enhanced time-frequency features;

[0013] Based on the one-dimensional wavelet transform convolution layer, one-dimensional wavelet transform convolution processing is performed on the multi-channel and multi-scale local enhanced time-frequency features to capture the temporal dependency in the multi-channel and multi-scale local enhanced time-frequency features and obtain the time-frequency features.

[0014] In some embodiments of the first aspect of the present application, identifying the voiceprint feature embedding vector to determine whether the current speaker is a registered user includes:

[0015] Identifying whether the voiceprint feature embedding vector corresponds to a registered voice vector of a registered user;

[0016] When the voiceprint feature embedding vector corresponds to a registered speech vector of a registered user, it is determined that the current speaker is a registered user;

[0017] When the voiceprint feature embedding vector does not correspond to the registered speech vectors of all registered users, it is determined that the current speaker is not a registered user.

[0018] In some embodiments of the first aspect of the present application, identifying whether the voiceprint feature embedding vector corresponds to a registered voice vector of a registered user includes:

[0019] Calculating the cosine similarity between the voiceprint feature embedding vector and at least one registered speech vector of the registered user;

[0020] averaging the cosine similarities of the registered speech vectors of the registered users to obtain recognition correspondence;

[0021] Based on the recognition correspondence, it is identified whether the voiceprint feature embedding vector corresponds to the registered voice vector of the registered user.

[0022] In some embodiments of the first aspect of the present application, identifying whether the voiceprint feature embedding vector corresponds to a registered voice vector of a registered user based on the recognition correspondence includes:

[0023] Determining whether the recognition correspondence is greater than a preset threshold;

[0024] If the recognition correspondence is greater than a preset threshold, the current speaker is determined to be a registered user;

[0025] If the recognition correspondence is less than or equal to a preset threshold, it is determined that the current speaker is not a registered user.

[0026] In some embodiments of the first aspect of the present application, the method for training the voiceprint recognition model includes:

[0027] Randomly cutting the obtained multiple audio files into specified lengths as multiple audio data, and adding random noise to the multiple audio data to obtain multiple recognition training data;

[0028] Labeling the voiceprint identity corresponding to each recognition training data, and obtaining the predicted voiceprint identity of the recognition training data through the voiceprint recognition model;

[0029] The difference between the predicted voiceprint identity and the voiceprint identity is obtained based on a preset loss function to train the voiceprint recognition model and obtain the trained voiceprint recognition model. The training method of the voiceprint recognition model includes:

[0030] Randomly cutting the obtained multiple audio files into specified lengths as multiple audio data, and adding random noise to the multiple audio data to obtain multiple recognition training data;

[0031] Labeling the voiceprint identity corresponding to each recognition training data, and obtaining the predicted voiceprint identity of the recognition training data through the voiceprint recognition model;

[0032] The difference between the predicted voiceprint identity and the voiceprint identity is obtained based on a preset loss function to train the voiceprint recognition model and obtain the trained voiceprint recognition model.

[0033] In some embodiments of the first aspect of the present application, the voiceprint recognition model includes a voiceprint feature vector extraction module and a recognition module; the voiceprint feature vector extraction module is used to perform feature extraction on the Mel-cepstral coefficient feature to obtain a voiceprint feature embedding vector; the recognition module is used to identify the voiceprint feature embedding vector to determine whether the current speaker is a registered user; wherein, the voiceprint feature vector extraction module includes a time-frequency feature extraction unit and a voiceprint feature vector generation unit; the time-frequency feature extraction unit includes a two-dimensional wavelet transform convolution layer, a two-dimensional residual network, a frequency attention network and a one-dimensional wavelet transform convolution layer; the voiceprint feature vector generation unit includes a global pooling layer.

[0034] A second aspect of the present application provides a voiceprint recognition system based on wavelet transform convolution, the system comprising:

[0035] An acquisition module is configured to acquire the speech to be recognized of the current speaker;

[0036] A first feature extraction module is configured to extract Mel-cepstral coefficient features based on the voiceprint characteristics of the speech to be recognized;

[0037] A second feature extraction module is configured to input the Mel-cepstral coefficient feature into a trained voiceprint recognition model to perform feature extraction on the Mel-cepstral coefficient feature to obtain a voiceprint feature embedding vector, including: performing multi-dimensional mixed wavelet transform convolution processing and frequency weight calibration processing on the Mel-cepstral coefficient feature to extract time-frequency features; converting frame-level features in the time-frequency features into sentence-level features to generate a voiceprint feature embedding vector; and

[0038] The voiceprint module is configured to identify the voiceprint feature embedding vector to determine whether the current speaker is a registered user.

[0039] In a third aspect, the present application provides a voiceprint recognition terminal, comprising: a processor and a memory; the memory is used to store a computer program; the processor is used to execute the computer program stored in the memory, so that the terminal executes any one of the wavelet transform convolution-based voiceprint recognition methods provided in the above embodiments.

[0040] A fourth aspect of the present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the voiceprint recognition method based on wavelet transform convolution provided in any one of the above embodiments is implemented.

[0041] As described above, this application provides a voiceprint recognition method, system, terminal, and medium based on wavelet transform convolution, which has the following beneficial effects:

[0042] This application introduces wavelet transform convolution into the voiceprint recognition task, which not only improves the model's ability to extract the time-frequency features of the speech signal, but also optimizes the design of the receptive field. The multi-resolution characteristics of the wavelet transform enable the model to capture the local and global features of the speech signal at different scales, thereby dynamically adjusting the range of the receptive field, so that it can pay attention to the time-frequency changes of details while taking into account the overall structure of the speech signal. In addition, wavelet transform convolution can effectively expand the receptive field, allowing the model to pay attention to a wider range of feature patterns, thereby more accurately extracting voiceprint information and improving the accuracy and robustness of recognition.

[0043] In the voiceprint recognition task, this application uses multi-dimensional mixed wavelet transform convolution to fully explore the time-frequency characteristics of the voice signal and improve the accuracy and robustness of feature extraction. Among them, the one-dimensional wavelet transform convolution mainly acts on the time dimension and can extract the instantaneous dynamic characteristics of the voice signal, while the two-dimensional wavelet transform convolution is processed in the time-frequency domain to more comprehensively capture the spatial relationship between different frequency components. By combining these two transformation methods, the model can focus on local details and global patterns at the same time, making the voiceprint features more stable and discriminative, thereby improving the recognition accuracy and noise resistance.

[0044] This application introduces a frequency attention mechanism into the voiceprint recognition task, enabling the model to autonomously focus on frequency information that is more critical to the recognition task. Through the attention mechanism, the model can dynamically analyze and weight the importance of different frequency regions, thereby more effectively extracting frequency domain information related to voiceprint features (such as formants, pitch, etc.) while suppressing interference from noise or irrelevant frequencies. This mechanism not only improves the model's ability to represent speech signals, but also enhances its robustness in complex environments, ultimately improving the accuracy and reliability of voiceprint recognition.

[0045] The voiceprint recognition model provided in this application not only provides high-precision recognition effects, but also has excellent lightweight characteristics, enabling it to perform well in resource-constrained environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Shown is a flow chart of a voiceprint recognition method based on wavelet transform convolution in one embodiment of the present application.

[0047] Figure 2 Shown is a structural diagram of a voiceprint recognition model in one embodiment of the present application.

[0048] Figure 3 Shown is a schematic diagram of the training and application of the voiceprint recognition model in one embodiment of the present application.

[0049] Figure 4 Shown is a structural diagram of a voiceprint recognition system based on wavelet transform convolution in one embodiment of the present application.

[0050] Figure 5 Shown is a structural diagram of a voiceprint recognition terminal in one embodiment of the present application. DETAILED DESCRIPTION

[0051] The following describes the embodiments of the present application through specific examples. Those skilled in the art can easily understand the other advantages and effects of the present application from the content disclosed in this specification. The present application can also be implemented or applied through other different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0052] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present application. Therefore, the illustrations only show components related to the present application and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0053] In addition, the descriptions of "first" and "second" in this application are for descriptive purposes only and should not be understood as indicating or implying their relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between the various embodiments can be combined with each other, but this must be based on the fact that they can be implemented by ordinary technicians in this field. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0054] Before further explaining the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations:

[0055] CNN: Convolutional Neural Network. A deep learning model designed specifically for processing grid-structured data (such as images and audio).

[0056] TDNN: Time Delay Neural Network, a special type of neural network architecture, is mainly used to process time series data or data with time delay characteristics.

[0057] MFCC: Mel-Frequency Cepstral Coefficients, a feature extraction method widely used in speech signal processing. It simulates the auditory characteristics of the human ear (Mel-scale filtering) to convert time-domain speech signals into low-dimensional cepstral feature vectors (usually 12-20 dimensions) that reflect the characteristics of the vocal tract, effectively capturing the timbre and formant information in the speech.

[0058] FBank: Filter Bank. The front-end features of MFCC, that is, the features before discrete cosine transform calculation.

[0059] FA stands for Frequency Attention. A technical concept used in signal processing, particularly in speech, audio, and image processing. The frequency attention mechanism focuses on the frequency domain characteristics of the data and uses an attention mechanism to highlight important frequency components while suppressing less important ones.

[0060] Res2Net: An improved residual network structure that constructs a hierarchical granular feature representation within a single residual block through hierarchical residual connections and grouped feature fusion mechanisms. This significantly improves the model's ability to capture multi-scale visual patterns and outperforms traditional residual network structures in tasks such as image classification and object detection.

[0061] AAM-Softmax: Additive Angular Margin Softmax. A loss function for classification tasks, commonly used in tasks such as speaker recognition, to enhance the model's ability to distinguish between class boundaries.

[0062] Voiceprint recognition: A technology that determines the identity of a speaker based on a voice signal, that is, determining which known speaker a voice comes from.

[0063] Voiceprint feature embedding vector: A feature vector generated by a model to describe a segment of speaker audio data and used to represent the speaker's information.

[0064] WT: Wavelet Transform. An effective signal analysis method that can provide multi-resolution local feature representation by decomposing the signal into time-frequency components of different scales.

[0065] WT-Conv: Wavelet Transform Convolution. It combines wavelet transform and convolution operations, performs multi-scale time-frequency decomposition of the signal through wavelet transform, and then extracts the local features of the signal through convolution.

[0066] Haar wavelet: Haar wavelet, the simplest orthogonal wavelet, consists of a set of step functions based on rectangular functions and is often used in signal processing, image compression and edge detection.

[0067] Daubechies wavelet: Daubechies wavelet is a type of wavelet with compact support and orthogonality. It has high regularity and vanishing moment and is suitable for signal denoising and feature extraction.

[0068] Coiflet wavelet: Coiflet wavelet is an improvement based on Daubechies wavelet. It optimizes the symmetry between the scaling function and the wavelet function while maintaining compact support and orthogonality. It is often used in high-precision signal reconstruction and image processing.

[0069] The embodiments of the present application provide a voiceprint recognition method, system, terminal and medium based on wavelet transform convolution. A hierarchical feature extraction mechanism is formed through a model structure based on multi-dimensional hybrid transform convolution and frequency domain attention mechanism. By combining cascaded wavelet transform layers with small-size convolution kernels, the frequency domain information of the convolution layer input features can be adaptively focused at different levels. While maintaining computational efficiency, the fine extraction and fusion of multi-band features are achieved, thereby enabling accurate voiceprint recognition with less computational effort.

[0070] Figure 1 The figure shows a flow chart of the voiceprint recognition method based on wavelet transform convolution provided by the embodiment of the present application. Figure 1 As shown, the voiceprint recognition method based on wavelet transform convolution provided in the embodiment of the present application includes steps S1 to S4.

[0071] S1. Obtain the speech to be recognized of the current speaker.

[0072] In some embodiments, the voiceprint recognition method based on wavelet transform convolution provided in the embodiments of the present application can be applied to various application terminals to obtain the voice to be recognized of the current user. For example, smart home devices installed at home can be controlled by voice control. When the user issues a command, the device first obtains the voice to be recognized of the current speaker through the sound collection unit, so as to subsequently identify whether the user's voiceprint is the voiceprint of a home user. Or when a user wants to use mobile banking, the mobile phone first obtains the user's voice to be recognized through the microphone, so as to subsequently identify whether the user's voiceprint is that of a bank user.

[0073] S2. Extract Mel-frequency cepstral coefficient features based on the voiceprint characteristics of the speech to be recognized.

[0074] In one embodiment, Mel-cepstral coefficient features related to human auditory characteristics are extracted from the acquired speech to be recognized. Because Mel-cepstral coefficient features are calculated in the frequency domain and the signal is compressed through logarithmic energy calculation, they are highly robust to a certain degree of noise. In some embodiments, extracting Mel-cepstral coefficient features based on the voiceprint characteristics of the speech to be recognized includes steps S21 to S26.

[0075] S21 : Pre-emphasize the speech to be recognized to generate corresponding high-frequency spectrum speech data.

[0076] The spectral characteristics of speech signals typically feature higher energy in low-frequency components and lower energy in high-frequency components. Pre-emphasis flattens the spectrum by increasing the energy of high-frequency components in speech signals. In practical applications, high-frequency components are often more susceptible to noise. Pre-emphasis enhances these high-frequency components, thereby improving the signal-to-noise ratio.

[0077] In some embodiments, the speech to be recognized is pre-emphasized by a first-order high-pass filter, and the transfer function of the filter is shown in formula (1).

[0078] y(t)=x(t)-αx(t-1) Formula (1)

[0079] Wherein, y(t) is the current sample of the high-frequency spectrum speech data generated after pre-emphasis; x(t) and x(t-1) are the current sample and the previous sample of the input speech data to be recognized, respectively; α is the filter coefficient. In a preferred embodiment, α takes a value between 0.9 and 1.

[0080] S22. Divide the high-frequency spectrum speech data into frames to generate a plurality of frames of speech data.

[0081] In most cases, speech signals are non-stationary, and it is meaningless to perform Fourier transform on the entire signal because the frequency contour of the signal will be lost over time. However, speech signals are short-term stationary signals. Therefore, Fourier transform is performed on short-time frames, and a good approximation of the signal frequency contour is obtained by connecting adjacent frames. After performing a framing operation on the high-frequency spectrum audio data, several frames of speech data are obtained. In one embodiment, the frame length of each frame of speech data is set to about 20-40ms. And in order to avoid excessive changes in the frame speech data of two adjacent frames, there is an overlapping area between the frame speech data of two adjacent frames, that is, a frame shift. In some embodiments, the frame shift is set to 10ms, but it should be noted that the present application is not limited to this.

[0082] S23 , performing windowing on each frame of speech data to generate a number of corresponding frames of windowed speech data.

[0083] After framing the data, each frame of speech data is multiplied by a window function to increase the continuity between the left and right ends of the frame, offset the assumption of data infinity in the subsequent fast Fourier transform, and reduce spectral leakage. In one embodiment, a Hamming window function can be used to perform the windowing operation, which can be expressed as Equation (2).

[0084] s(t)=W(n,a)×v(t) Formula (2)

[0085] Among them, v(t) is the frame speech data before windowing, s(t) is the frame windowed speech data after windowing, and W(n,a) is the Hamming window function.

[0086] Furthermore, the mathematical expression of the Hamming window function is shown in formula (3).

[0087] W(n,a)=1-aa×cos(2πn) / (N-1) Formula (3)

[0088] Where, 0≤n≤N-1, N is the window length, that is, the number of sample points; a is the coefficient of the window function.

[0089] S24. Perform fast Fourier transform on each frame of windowed speech data to generate linear spectrum data.

[0090] Since it is usually difficult to see the characteristics of the signal from its changes in the time domain, it is necessary to perform a fast Fourier transform on the signal to convert it into an energy distribution in the frequency domain. In this way, different speech characteristics can be identified by observing the energy distribution in the frequency domain.

[0091] In some embodiments, the calculation formula for performing fast Fourier transform on each frame of windowed speech data is shown in formula (4).

[0092]

[0093] Wherein, 0≤k≤N-1, N is the number of calculation points of the fast Fourier transform, which is set to 256 or 512 in a preferred embodiment; s(n) is the time domain signal, and S(k) is the converted linear spectrum data, representing the frequency domain signal.

[0094] Then the power spectrum P of the frequency domain signal data S(k) is calculated according to formula (5).

[0095]

[0096] S25 . Perform Mel filtering on the linear spectrum data to generate Mel filtering signal data.

[0097] In some embodiments, a Mel-scale filter bank is applied to the power spectrum of a signal to extract frequency bands. The Mel filter bank simulates the human auditory perception system, that is, the human ear, which only pays attention to certain specific frequency components. Therefore, the Mel filter bank has different sensitivities to signals of different frequencies, that is, it only allows signals of certain frequencies to pass through, while filtering out those frequency signals that you do not want to perceive. However, the distribution of these filters on the frequency coordinate axis is not uniform. There are many filters in the low-frequency region, and the distribution is relatively dense; in the high-frequency region, the number of filters is small, and the distribution is sparse. Therefore, the purpose of the Mel filter bank is to simulate the nonlinear perception of sound by the human ear, and to have stronger discrimination at lower frequencies.

[0098] In some embodiments, performing Mel filtering on the linear spectrum data includes:

[0099] First, determine the Mel frequency range to be covered. For speech data, the Mel frequency range can be from low frequency 0 to high frequency.

[0100] The calculation formula of Mel frequency is shown in formula (6).

[0101]

[0102] Where, f is the physical frequency; f mel is the Mel frequency.

[0103] Next, M+2 points are uniformly selected within the Mel frequency range as Mel frequency points, m(0), m(1), …, m(M+1). Here, M is the number of filters. The Mel frequency points determine the center frequency of the filter. These Mel frequency points are then converted back to physical frequencies using the inverse Mel frequency conversion formula (e.g., formula (7)).

[0104]

[0105] Finally, a triangular filter is constructed to complete the Mel filtering operation. The calculation formula is shown in formula (8).

[0106]

[0107] Among them, the Mel filter bank includes M triangular filters, H m (k ′ ) is the frequency response of the m-th triangular filter, k ′ represents the kth ′ The center frequency of the mth triangular filter is f(m), the response at the center frequency is 1, and decreases linearly toward 0 until the center frequency responses of the two adjacent filters are 0; the interval between the center frequencies f(m) widens as the value of m increases.

[0108] S26. Perform a logarithmic operation on the Mel filter signal data to obtain Mel cepstral coefficient features.

[0109] In order to improve the robustness of the Mel spectrum data to a certain degree of noise, it is necessary to compress each Mel spectrum data through logarithmic energy calculation. The specific calculation formula is shown in formula (9).

[0110] FBanks=20×log 10 (H m (k')×P) Formula (9)

[0111] Among them, H m (k') is the frequency response of the mth triangular filter; P is the power spectrum of the linear spectrum data; and FBanks is the Mel-frequency cepstral coefficient feature. It can be seen that the Mel-frequency cepstral coefficient feature is actually a single-channel two-dimensional feature data of 1×N×F. Its height N is the number of Mel-frequency filters used when performing the Mel-frequency filtering operation, and its width F is the number of frames when performing the framing operation.

[0112] S3. Input the Mel-cepstral coefficient feature into a trained voiceprint recognition model to perform feature extraction on the Mel-cepstral coefficient feature to obtain a voiceprint feature embedding vector.

[0113] In some embodiments, feature extraction is performed on the Mel-cepstral coefficient features to obtain a voiceprint feature embedding vector, including: performing multi-dimensional mixed wavelet transform convolution processing and frequency weight calibration processing on the Mel-cepstral coefficient features to extract time-frequency features; converting the frame-level features in the time-frequency features into sentence-level features to generate a voiceprint feature embedding vector.

[0114] Figure 2 Shown is a schematic diagram of the structure of the voiceprint recognition model provided in the embodiment of this application. Figure 2 As shown, the voiceprint recognition model includes a voiceprint feature vector extraction module, which is used to extract Mel-frequency cepstral coefficient features and obtain a voiceprint feature embedding vector.

[0115] Among them, continue to refer to Figure 2 The voiceprint feature vector extraction module includes a time-frequency feature extraction unit and a voiceprint feature vector generation unit. The time-frequency feature extraction unit is used to perform multi-dimensional mixed wavelet transform convolution processing and frequency weight calibration on the Mel-frequency cepstral coefficient features to extract time-frequency features. The voiceprint feature vector generation unit is used to convert the frame-level features in the time-frequency features into sentence-level features to generate a voiceprint feature embedding vector.

[0116] Continue to refer Figure 2The time-frequency feature extraction unit further includes a one-dimensional wavelet transform convolution layer, a two-dimensional wavelet transform convolution layer, a two-dimensional residual network, and a frequency attention network. Wavelet transform convolution introduces the multi-scale characteristics of the wavelet transform into traditional convolutional neural networks to enhance the ability to capture signal features across different frequency ranges and scales. Traditional convolution extracts local features of the input signal by learning a convolution kernel, while wavelet transform convolution combines the convolution kernel with wavelet basis functions, allowing the convolution operation to simultaneously consider features from different frequency ranges. Wavelet functions of different scales can capture information from different frequency bands of the signal, which is very effective for processing signals with multi-layered structures. By performing multi-scale wavelet transforms on the signal, local features at different frequencies can be extracted. These frequency band features are then fused through convolution. Furthermore, the convolution kernel not only operates on local regions in the time or frequency domain but also adapts to the features of different frequencies, thereby improving the flexibility and robustness of feature extraction.

[0117] Compared to traditional convolutional neural networks, wavelet transform convolution can significantly expand the receptive field (i.e., the range of dependence on input data) at a smaller convolution kernel size. The multi-scale nature of the wavelet transform enables the network to capture a wider range of features at a lower computational cost, which is particularly suitable for deploying voiceprint recognition algorithms on embedded devices, especially for tasks that require capturing long-term dependencies between input features. Therefore, the time-frequency feature extraction unit can fully exploit the time-frequency characteristics of the speech signal through multi-dimensional mixed wavelet transform convolution processing, improving the accuracy and robustness of feature extraction.

[0118] At the same time, the time-frequency feature extraction unit uses a two-dimensional residual network for group convolution, which not only preserves the multi-scale feature expression but also significantly reduces the computational complexity and number of parameters. In other words, the two-dimensional residual network not only more flexibly captures the relationship between detailed features and global information, but also significantly improves the efficiency and performance of the model under limited computing resources. Therefore, the two-dimensional residual network optimizes the computational cost while enhancing the model's expressiveness.

[0119] In addition, the time-frequency feature extraction unit also introduces a frequency attention network, which can effectively model features in different frequency ranges and enable the model to adaptively focus on more important frequency information.

[0120] In some embodiments, extracting time-frequency features includes steps S311 to S314.

[0121] S311 . Perform two-dimensional wavelet transform convolution processing on the Mel-cepstral coefficient feature based on a two-dimensional wavelet transform convolution layer to perform multi-scale time-frequency decomposition on the Mel-cepstral coefficient feature to obtain multi-scale local time-frequency features.

[0122] Two-dimensional wavelet transform convolution is used to extract local time-frequency feature information and model the importance of features in different frequency ranges. Specifically, two-dimensional wavelet transform convolution with a small receptive field extracts local time and frequency information. By utilizing repeated local frequency features with small frequency offsets, the model can effectively model finer frequency details with fewer feature channels.

[0123] In some embodiments, the two-dimensional wavelet transform convolution layer includes a wavelet transform layer, a convolution layer, and an inverse wavelet transform reconstruction layer. The calculation process of the wavelet transform convolution can be expressed by formula (10).

[0124] Z=IWT(Conv(W,WT(X)) Formula (10)

[0125] Among them, X is the input Mel-frequency cepstral coefficient feature, Z is the output multi-scale local time-frequency feature, WT and IWT are wavelet transform and inverse wavelet transform respectively, Conv is the standard two-dimensional convolution operator, and W is the parameter of the two-dimensional convolution operator.

[0126] In some embodiments, when starting the wavelet transform convolution, a wavelet basis must be selected first. Preferably, the Haar wavelet is selected as the wavelet basis for the two-dimensional wavelet transform convolution. The filter operator of the two-dimensional Haar is shown in formula (11) and formula (12):

[0127]

[0128] Among them, f LL is a low frequency filter, f LH ,f HL ,f HH is a high frequency filter bank.

[0129] It should be noted that the wavelet basis in wavelet transform convolution is not limited to the Haar wavelet basis. Other types of wavelet basis can be selected according to task requirements, such as Daubechies wavelet, Coiflet wavelet, etc., to better adapt to different feature extraction requirements.

[0130] Afterwards, a multi-level wavelet transform layer is used to perform a two-dimensional wavelet transform on each channel of the input Mel-frequency cepstral coefficient feature to obtain the decomposition features of different frequency bands. The process is shown in formula (13).

[0131]

[0132] Where i represents the i-th level wavelet transform, 1≤i≤A, and A is the level of wavelet transform; is the low-frequency approximate component after decomposition by the i-th wavelet transform layer, is the low-frequency approximate component of the i-1th level wavelet transform layer, and serves as the input of the i-th level wavelet transform layer. is the high-frequency detail component after decomposition of the i-th wavelet transform layer; is the input of the entire wavelet transform layer.

[0133] After the wavelet transform, the convolution layer is used to apply the convolution operation on each wavelet transform layer to extract features. The convolution kernel is combined with the multi-scale characteristics of the wavelet transform to further capture the features of different frequency bands. The process is shown in formula (14).

[0134]

[0135] in, for The output after applying the convolution operation, for The output after applying the convolution operation, W (i) is the convolution weight of the i-th wavelet transform layer.

[0136] Afterwards, the inverse wavelet transform reconstruction layer is used to combine the features of different frequency bands decomposed above to reconstruct the output features. The process is shown in formula (15).

[0137]

[0138] where Z (i) is the reconstructed feature of the i-th level inverse wavelet transform, that is, the multi-scale local time-frequency feature.

[0139] S312. Grouping multiple feature channels of the multi-scale local time-frequency features based on a two-dimensional residual network, and performing convolution processing on the multi-scale local time-frequency features of each group of feature channels to obtain multi-channel multi-scale local time-frequency features.

[0140] In some embodiments, the two-dimensional residual network includes multiple two-dimensional convolutional layers, which are used to group multiple feature channels of multi-scale local time-frequency features and perform convolution processing on the multi-scale local time-frequency features of each group of feature channels, thereby giving the voiceprint recognition model the ability to extract time-frequency features at different scales. The convolution is followed by batch normalization (BN) and ReLU activation function. The input is dimensionalized through 1×1 convolution and then added to the convolution result (residual connection). The ReLU activation function is then used to output multi-channel and multi-scale local time-frequency features.

[0141] S313. Based on the frequency attention network, the frequency dimension of the multi-channel and multi-scale local time-frequency features is weighted and calibrated to obtain the multi-channel and multi-scale local enhanced time-frequency features.

[0142] By applying the frequency attention mechanism on the frequency dimension of the multi-channel and multi-scale local time-frequency features, the time-frequency feature extraction unit can accurately capture the relatively important frequency components in the local time domain features, thereby enhancing its ability to focus on key information. In some embodiments, given a multi-channel and multi-scale local time-frequency feature X of shape [C, M, N] (for the sake of brevity, the batch dimension is omitted here, where C represents the number of feature channels, M represents the frequency dimension, and N represents the time dimension), the weight distribution on the frequency dimension is calculated using the following formulas (16) and (17).

[0143]

[0144] Q=σ(W2δ(W1F+b1)+b2) Formula (17)

[0145] Among them, the input multi-channel multi-scale local time-frequency feature X outputs the single-channel multi-scale local time-frequency feature F of [1, M, N] through formula (16), that is, f jk ; W1, W2, b1, b2 are learnable parameters, such as convolution weights and biases; σ is the sigmoid function; δ is the activation function, such as the ReLU function; Q is the weight of the frequency dimension obtained by paying attention to the single-channel multi-scale local time-frequency feature F, and its shape is [1, M, N].

[0146] This weight Q is then multiplied by the input multi-channel and multi-scale local time-frequency feature X to obtain the multi-channel and multi-scale local enhanced time-frequency feature Y after frequency recalibration, whose dimension is the same as the input feature:

[0147] Y=Q·X

[0148] This process not only helps the model automatically identify and emphasize important features in different frequency ranges, but also makes the model's expression of time-frequency features more detailed, thereby improving the overall representation ability and performance.

[0149] S314. Perform one-dimensional wavelet transform convolution processing on the multi-channel and multi-scale local enhanced time-frequency features based on the one-dimensional wavelet transform convolution layer to capture the temporal dependency in the multi-channel and multi-scale local enhanced time-frequency features and obtain the time-frequency features.

[0150] Referring to the content of the aforementioned two-dimensional wavelet transform convolution, in fact, the one-dimensional wavelet transform convolution also combines the one-dimensional convolution and wavelet transform strategies, expanding the receptive field of the model, enabling it to capture a wider range of sequence information, thereby efficiently modeling the temporal dependencies in the sequence data. This structure not only improves the accuracy of the model when processing dynamic data, but also optimizes the computational efficiency. In some embodiments, the Haar wavelet is selected as the wavelet basis of the one-dimensional wavelet transform convolution. Among them, the filter operator of the one-dimensional Haar is shown in formula (18):

[0151]

[0152] Among them, f L is a low frequency filter, f H For high frequency filter.

[0153] Furthermore, the one-dimensional wavelet transform convolution layer can also cooperate with the aforementioned two-dimensional residual network to form a powerful feature extraction capability. One-dimensional convolution hard-codes the absolute frequency position of the feature through the order of the filter coefficients, which is crucial for voiceprint recognition tasks because one-dimensional convolution can capture the unique frequency patterns in the speaker's voice. The aforementioned two-dimensional residual network further enhances the model's modeling capabilities at multiple scales. By dividing the feature channels into different groups, the model allows the model to model the temporal relationship of sequence data at different scales. This combination not only expands the receptive field of the feature, allowing the model to pay attention to local and global temporal information simultaneously, but also can extract important features from a richer frequency range.

[0154] Continue to refer Figure 2 In some embodiments, the voiceprint feature vector generation unit includes a global pooling layer. The frame-level features are converted into sentence-level features through the global pooling layer, thereby providing a global representation of the entire speech segment. Specifically, frame-level features are extracted from each frame of the input speech signal, and these features usually contain local information at each time step. However, the voiceprint recognition task not only needs to capture local temporal features, but also needs to integrate these features globally to more comprehensively characterize the characteristics of the entire sentence (speech segment). Therefore, through a global pooling layer, such as average pooling or maximum pooling, the information in the time dimension can be compressed into a sentence-level vector of fixed length by performing aggregation operations on all frame-level features, which is used as the voiceprint feature embedding vector.

[0155] S4. Identify the voiceprint feature embedding vector to determine whether the current speaker is a registered user.

[0156] Continue to refer Figure 2The voiceprint recognition model also includes a recognition module, which is used to identify the voiceprint feature embedding vector to determine whether the current speaker is a registered user. In some embodiments, the recognition module includes a linear layer, which is used to identify whether the voiceprint feature embedding vector corresponds to the registered voice vector of a registered user; when the voiceprint feature embedding vector corresponds to the registered voice vector of a registered user, the current speaker is determined to be a registered user; when the voiceprint feature embedding vector does not correspond to the registered voice vectors of all registered users, the current speaker is determined to be a non-registered user. In other words, the recognition module determines whether the current speaker is a registered user by determining whether the voiceprint feature embedding vector corresponds to the registered voice vector.

[0157] In some embodiments, identifying whether the voiceprint feature embedding vector corresponds to the registered voice vector of the registered user includes: calculating the cosine similarity between the voiceprint feature embedding vector and at least one registered voice vector of the registered user; performing weighted calculation on the cosine similarities of each registered voice vector of the registered user to obtain recognition correspondence; and identifying whether the voiceprint feature embedding vector corresponds to the registered voice vector of the registered user based on the recognition correspondence.

[0158] That is, the recognition module actually uses cosine similarity to calculate the similarity between the voiceprint feature embedding vector and each registered voice vector of each registered user in the registered voiceprint feature embedding vector database, and averages the cosine similarity of each registered voice vector of each registered user to obtain the recognition correspondence. Based on the recognition correspondence, it is determined whether the voiceprint feature embedding vector corresponds to the registered voice vector of a registered user. The specific process of obtaining the recognition correspondence is shown in formula (19).

[0159]

[0160] in, Indicates the calculation of vector x and The cosine similarity of , where x is the voiceprint feature embedding vector of the speech to be recognized; The jth segment registered speech vector for the wth registered user; s i is the similarity between the voiceprint feature embedding vector of the speech to be recognized and the i-th registered user, that is, the recognition correspondence, N is the number of registered users, and M is the number of voice segments registered by each registered user. In some embodiments, M = 3, but it should be noted that this application is not limited to this.

[0161] In some embodiments, the method for obtaining the registered voiceprint feature embedding vector database includes: extracting the Mel-cepstral coefficient features of the voice data of each registered user; inputting the Mel-cepstral coefficient features into the voiceprint recognition model to generate each registered voice vector corresponding to each registered user through the voiceprint feature vector extraction module of the voiceprint recognition model, and storing the registered users and the corresponding registered voice vectors in the registered voiceprint feature embedding vector database.

[0162] Therefore, this application introduces a weighted strategy for the similarity of multiple registered voice segments for each registered user, effectively alleviating the problem of reduced recognition accuracy caused by a single registered voice environment or a low signal-to-noise ratio. Furthermore, by weighting the similarity of multiple registered voice segments, the voiceprint recognition model can more comprehensively capture the diversity of user voiceprint features, thereby improving its adaptability to complex environments and noise, and further enhancing the model's robustness and recognition accuracy.

[0163] It should be noted that when calculating the recognition correspondence, a variety of strategies can be adopted, not limited to averaging the similarities of multiple registered voice segments. For example, you can choose to assign weights to different registered voice vectors to perform a weighted calculation of the cosine similarity of each registered voice vector to ensure that registered voices with more reference value have higher weights; or you can choose the maximum or median of the similarities of multiple voice segments as the final score to enhance robustness against outliers; or use machine learning methods (such as classifiers or regression models) to conduct a comprehensive analysis of the similarities of multiple voice segments, thereby more flexibly capturing the diversity and complexity of user voiceprint features. These strategies can optimize recognition performance according to specific scenarios and needs, and improve the adaptability and accuracy of the model.

[0164] Furthermore, determining whether the voiceprint feature embedding vector corresponds to the registered user's registered voice vector based on the recognition correspondence includes: determining whether the recognition correspondence is greater than a preset threshold; if the recognition correspondence is greater than the preset threshold, determining that the current speaker is a registered user; if the recognition correspondence is less than or equal to the preset threshold, determining that the current speaker is not a registered user. The preset threshold for the recognition correspondence can be set and adjusted by the user as needed, and this application is not limited thereto.

[0165] In practice, cosine similarity measures the degree of similarity in the auditory features of two speech segments by measuring the angular difference between two vectors in a high-dimensional space. A higher similarity indicates a greater likelihood that the two segments originate from the same user; a lower similarity suggests that the two segments likely originate from different users. This process provides an accurate and efficient basis for voiceprint recognition. Therefore, setting an appropriate preset threshold also facilitates accurate voiceprint recognition.

[0166] In summary, the present embodiment proposes a voiceprint recognition model that significantly improves its ability to capture different frequency features by organically combining multi-dimensional hybrid wavelet transform convolution, an attention mechanism, and a Res2Net network structure. Furthermore, a multi-segment speech similarity weighting strategy is employed to effectively enhance the recognition accuracy of voiceprint features.

[0167] In some embodiments, the training method of the voiceprint recognition model includes: randomly cutting the obtained multiple audio files into specified lengths as multiple audio data, and adding random noise to the multiple audio data to obtain multiple recognition training data; labeling the voiceprint identity corresponding to each recognition training data, and obtaining the predicted voiceprint identity of the recognition training data through the voiceprint recognition model; obtaining the difference between the predicted voiceprint identity and the voiceprint identity based on a preset loss function to train the voiceprint recognition model and obtain a trained voiceprint recognition model.

[0168] In some embodiments, the present application uses the development set (dev) of VoxCeleb1 and the test set (test) of VoxCeleb2, both public datasets in the field of voiceprint recognition, where the former is used for model training and the latter is used for model testing. The speech in both datasets is in English, and there is no overlap in users, that is, users in VoxCeleb1 do not exist in VoxCeleb2, and vice versa. Furthermore, for the training set, the present application randomly cuts speech segments of a fixed length (such as 3 seconds) as input for voiceprint recognition model training. For the test set, there is no restriction on the audio length, and the model inference stage can accept input audio of any length.

[0169] It should be noted that for the training set, if the original audio file is not long enough, it is padded symmetrically to the required length as the model input.

[0170] In addition, in order to improve the environmental robustness of the model in actual use, this application adds random noise (such as environmental noise, echo, etc.) to the input audio with a certain probability (such as 50%) during the training process as recognition training data.

[0171] Afterwards, the recognition training data is annotated with the real user's voiceprint identity, that is, the user identity corresponding to the recognition training data is annotated. Then, the predicted voiceprint identity of the recognition training data is obtained through the voiceprint recognition model. Finally, the difference between the predicted voiceprint identity and the voiceprint identity is obtained based on a preset loss function to train the voiceprint recognition model.

[0172] In some embodiments, the AAM-Softmax function is used as the loss function, and the calculation formula is shown in formula (20).

[0173]

[0174] Among them, L AAM-Softmax is the loss value of the AAM-Softmax loss function; θ y is the cosine of the angle between the embedding feature vector and the weight vector of the target category y, θ j is the cosine of the angle between the embedded feature vector and the weight vector of the non-target category j; m is a hyperparameter, or called the angle interval; s is a scaling factor used to control the stability of the value and the gradient change.

[0175] In fact, the AAM-Softmax loss function enhances the model's ability to distinguish in the feature space by introducing angle intervals and normalization processing, especially in dealing with highly similar categories, difficult samples, and tasks that require strong generalization capabilities.

[0176] Furthermore, during parameter optimization, the Adam optimizer, featuring an adaptive learning rate optimization algorithm, dynamically adjusts the learning rate of each parameter, accelerating model convergence. To further improve model training effectiveness, a periodic learning rate adjustment strategy is employed. This periodic learning rate adjustment strategy allows for targeted increases or decreases in the learning rate at different stages of model training, promoting rapid convergence in the early stages, effectively avoiding local optimal solutions in the later stages, and enhancing the model's generalization and stability during training.

[0177] The voiceprint recognition model trained using the above example achieved excellent performance on the test set of the public dataset VoxCeleb1, achieving an excellent equal error rate (EER) of 2.33%. This demonstrates that the accuracy and robustness of the model in speaker recognition tasks have been greatly improved. At the same time, the number of model parameters is controlled at the K level, demonstrating extremely high parameter efficiency. This compact model design not only provides high-precision recognition results, but also has excellent lightweight characteristics, enabling it to perform well even in resource-constrained environments.

[0178] Figure 3 The following is a diagram showing the training and application of the voiceprint recognition model provided in the embodiment of the present application. Figure 3As shown, the voiceprint recognition model is first trained using the training speech as a data set, including data processing of the training speech, such as randomly cropping speech segments of a fixed length (such as 3 seconds), and extracting Mel-frequency cepstral coefficient features based on this as input for voiceprint recognition model training. The voiceprint feature embedding vector is then extracted according to the voiceprint feature vector extraction module, and the classification module is used to perform category prediction operations, mapping the voiceprint feature embedding vector extracted from the audio to a specific speaker category. The classification results are evaluated through a loss function, thereby continuously updating the model parameters during the training process. When the voiceprint recognition model training is completed, it is necessary to obtain a database of registered voiceprint feature embedding vectors. After that, the registered database can be used for voiceprint recognition, and the cosine similarity between the voiceprint feature embedding vector extracted from the speech to be recognized input by the current speaker and the registered voice of the registered user in the database is obtained to obtain the recognition correspondence, and then determine whether the current speaker is a registered user.

[0179] It should be noted that the above-mentioned classification module is limited to the model training process. That is, during the training phase, the voiceprint recognition model uses the classification module to adapt to different sound characteristics and environmental conditions, thereby improving the accuracy and robustness of the voiceprint feature embedding vector extraction in the previous step. After the training is completed, this application divides the trained voiceprint recognition model into a voiceprint feature vector extraction module and a recognition module, where the recognition module is used to identify the voiceprint feature embedding vector to determine whether the current speaker is a registered user.

[0180] Figure 4 The following is a schematic diagram of the structure of the voiceprint recognition system based on wavelet transform convolution provided by the embodiment of the present application. Figure 4 As shown, the voiceprint recognition system 40 based on wavelet transform convolution includes an acquisition module 401 , a first feature extraction module 402 , a second feature extraction module 403 and a voiceprint module 404 .

[0181] The acquisition module 401 is configured to acquire the speech to be recognized of the current speaker.

[0182] The first feature extraction module 402 is configured to extract Mel-cepstral coefficient features based on the voiceprint characteristics of the speech to be recognized.

[0183] The second feature extraction module 403 is configured to input the Mel-cepstral coefficient features into the trained voiceprint recognition model to perform feature extraction on the Mel-cepstral coefficient features and obtain a voiceprint feature embedding vector, including: performing multi-dimensional mixed wavelet transform convolution processing and frequency weight calibration processing on the Mel-cepstral coefficient features to extract time-frequency features; converting the frame-level features in the time-frequency features into sentence-level features to generate a voiceprint feature embedding vector.

[0184] The voiceprint module 404 is configured to identify the voiceprint feature embedding vector to determine whether the current speaker is a registered user.

[0185] It should be noted that the principles of the acquisition module 401, the first feature extraction module 402, the second feature extraction module 403 and the voiceprint module 404 provided in the embodiment of the present application correspond one-to-one to the steps in the above-mentioned voiceprint recognition method based on wavelet transform convolution, so they will not be repeated here.

[0186] It should also be understood that the division of modules in the embodiments of the present application is illustrative and is merely a logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in the various embodiments of the present application may be integrated into a single processor, or may exist physically separately, or two or more modules may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or software functional modules.

[0187] The voiceprint recognition method based on wavelet transform convolution provided in the embodiment of the present application can be implemented on the terminal side or the server side. As for the hardware structure of the voiceprint recognition terminal, please refer to Figure 5 , is a schematic diagram of an optional hardware structure of a voiceprint recognition terminal 700 provided in an embodiment of the present application. The voiceprint recognition terminal 700 can be a mobile phone, a computer, a tablet device, a personal digital assistant, a factory backend processing device, etc. The voiceprint recognition terminal 700 includes: at least one processor 701, a memory 702, at least one network interface 704, and a user interface 706. Furthermore, the various components in the voiceprint recognition terminal 700 are coupled together via a bus system. It will be understood that the bus system is used to enable connection and communication between these components. In addition to including a data bus, the bus system also includes a power bus, a control bus, and a status signal bus. The user interface 706 can include a display, a keyboard, a mouse, a trackball, a click gun, keys, buttons, a touchpad, or a touch screen.

[0188] It can be understood that the memory 702 can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. This application does not specifically limit this. The memory 702 in the embodiment of the present application is used to store various categories of data to support the operation of the voiceprint recognition terminal 700. Examples of such data include: any executable program for operating on the voiceprint recognition terminal 700, such as an operating system 7021 and an application 7022; the operating system 7021 includes various system programs, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks. The application 7022 can include various applications, such as a media player (MediaPlayer), a browser (Browser), etc. The implementation of the voiceprint recognition method based on wavelet transform convolution provided in the embodiment of the present application can be included in the application 7022.

[0189] The voiceprint recognition method disclosed in the above embodiment of the present application can be applied to the processor 701, or implemented by the processor 701. The processor 701 may be an integrated circuit chip with signal processing capabilities. During implementation, the various steps of the above method can be completed by the hardware integrated logic circuit in the processor 701 or by instructions in the form of software. The above processor 701 can be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 701 can implement or execute the various methods, steps and logic block diagrams disclosed in the embodiments of the present application. The general-purpose processor 701 can be a microprocessor or any conventional processor, etc.

[0190] In an exemplary embodiment, the voiceprint recognition terminal 700 can be used by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), and complex programmable logic devices (CPLDs) to execute the aforementioned voiceprint recognition method based on wavelet transform convolution.

[0191] Those skilled in the art will appreciate that all or part of the steps in the above-described method embodiments can be implemented using hardware associated with a computer program. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps in the above-described method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0192] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

[0193] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Anyone skilled in the art may modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by one of ordinary skill in the art without departing from the spirit and technical concepts disclosed in this application shall be covered by the claims of this application.

Claims

1. A voiceprint recognition method based on wavelet transform convolution, characterized in that: The method comprises: Get the speech to be recognized of the current speaker; Extracting Mel-frequency cepstral coefficient features based on the voiceprint characteristics of the speech to be recognized; Inputting the Mel-cepstral coefficient features into a trained voiceprint recognition model to perform feature extraction on the Mel-cepstral coefficient features to obtain a voiceprint feature embedding vector, including: performing multi-dimensional mixed wavelet transform convolution processing and frequency weight calibration processing on the Mel-cepstral coefficient features to extract time-frequency features; converting frame-level features in the time-frequency features into sentence-level features to generate a voiceprint feature embedding vector; and The voiceprint feature embedding vector is identified to determine whether the current speaker is a registered user.

2. The voiceprint recognition method based on wavelet transform convolution according to claim 1 is characterized in that: The extracting of time-frequency features comprises: Performing a two-dimensional wavelet transform convolution process on the Mel-cepstral coefficient feature based on a two-dimensional wavelet transform convolution layer to perform multi-scale time-frequency decomposition on the Mel-cepstral coefficient feature to obtain multi-scale local time-frequency features; Grouping the multiple feature channels of the multi-scale local time-frequency features based on a two-dimensional residual network, and performing convolution processing on the multi-scale local time-frequency features of each group of feature channels to obtain multi-channel multi-scale local time-frequency features; Based on the frequency attention network, a weight calibration process is performed on the frequency dimension of the multi-channel and multi-scale local time-frequency features to obtain multi-channel and multi-scale local enhanced time-frequency features; Based on the one-dimensional wavelet transform convolution layer, one-dimensional wavelet transform convolution processing is performed on the multi-channel and multi-scale local enhanced time-frequency features to capture the temporal dependency in the multi-channel and multi-scale local enhanced time-frequency features and obtain the time-frequency features.

3. The voiceprint recognition method based on wavelet transform convolution according to claim 1 is characterized in that: Identifying the voiceprint feature embedding vector to determine whether the current speaker is a registered user includes: Identifying whether the voiceprint feature embedding vector corresponds to a registered voice vector of a registered user; When the voiceprint feature embedding vector corresponds to a registered speech vector of a registered user, it is determined that the current speaker is a registered user; When the voiceprint feature embedding vector does not correspond to the registered speech vectors of all registered users, it is determined that the current speaker is not a registered user.

4. The voiceprint recognition method based on wavelet transform convolution according to claim 3 is characterized in that: Identifying whether the voiceprint feature embedding vector corresponds to a registered voice vector of a registered user includes: Calculating the cosine similarity between the voiceprint feature embedding vector and at least one registered speech vector of the registered user; averaging the cosine similarities of the registered speech vectors of the registered users to obtain recognition correspondence; Based on the recognition correspondence, it is identified whether the voiceprint feature embedding vector corresponds to the registered voice vector of the registered user.

5. The voiceprint recognition method based on wavelet transform convolution according to claim 4 is characterized in that: Identifying whether the voiceprint feature embedding vector corresponds to the registered voice vector of the registered user based on the recognition correspondence includes: Determining whether the recognition correspondence is greater than a preset threshold; If the recognition correspondence is greater than a preset threshold, the current speaker is determined to be a registered user; If the recognition correspondence is less than or equal to a preset threshold, it is determined that the current speaker is not a registered user.

6. The voiceprint recognition method based on wavelet transform convolution according to claim 1 is characterized in that: The training method of the voiceprint recognition model includes: Randomly cutting the obtained multiple audio files into specified lengths as multiple audio data, and adding random noise to the multiple audio data to obtain multiple recognition training data; Labeling the voiceprint identity corresponding to each recognition training data, and obtaining the predicted voiceprint identity of the recognition training data through the voiceprint recognition model; The difference between the predicted voiceprint identity and the voiceprint identity is obtained based on a preset loss function to train the voiceprint recognition model and obtain the trained voiceprint recognition model.

7. The voiceprint recognition method based on wavelet transform convolution according to claim 1 is characterized in that: The voiceprint recognition model includes a voiceprint feature vector extraction module and a recognition module; the voiceprint feature vector extraction module is used to extract the Mel cepstral coefficient feature to obtain a voiceprint feature embedding vector; the recognition module is used to identify the voiceprint feature embedding vector to determine whether the current speaker is a registered user; wherein, The voiceprint feature vector extraction module includes a time-frequency feature extraction unit and a voiceprint feature vector generation unit; the time-frequency feature extraction unit includes a two-dimensional wavelet transform convolution layer, a two-dimensional residual network, a frequency attention network and a one-dimensional wavelet transform convolution layer; the voiceprint feature vector generation unit includes a global pooling layer.

8. A voiceprint recognition system based on wavelet transform convolution, characterized in that: The system comprises: An acquisition module, configured to acquire the speech to be recognized of the current speaker; A first feature extraction module is configured to extract Mel-cepstral coefficient features based on the voiceprint characteristics of the speech to be recognized; A second feature extraction module is configured to input the Mel-cepstral coefficient feature into a trained voiceprint recognition model to perform feature extraction on the Mel-cepstral coefficient feature to obtain a voiceprint feature embedding vector, including: performing multi-dimensional mixed wavelet transform convolution processing and frequency weight calibration processing on the Mel-cepstral coefficient feature to extract time-frequency features; converting frame-level features in the time-frequency features into sentence-level features to generate a voiceprint feature embedding vector; and The voiceprint module is configured to identify the voiceprint feature embedding vector to determine whether the current speaker is a registered user.

9. A voiceprint recognition terminal, characterized in that: include: processor and memory; The memory is used to store computer programs; The processor is configured to execute the computer program stored in the memory, so that the terminal executes the voiceprint recognition method based on wavelet transform convolution as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the voiceprint recognition method based on wavelet transform convolution according to any one of claims 1 to 7 is implemented.

Images