Feature-based intelligent voiceprint recognition method and system

Abstract

The invention belongs to the technical field of voiceprint recognition, and discloses an intelligent voiceprint recognition method and system based on features. Comprising the following steps: collecting a plurality of voiceprint data samples, and preprocessing each voiceprint data sample to obtain a comprehensive voice sample; constructing a voice directed graph model based on the comprehensive voice samples, and recognizing a core voice sample set from the voice directed graph model; performing feature extraction on the core voice sample set to obtain # imgabs0 # comprehensive feature vectors; obtaining a sample selection strategy by using a clustering algorithm based on the comprehensive feature vector and the core voice sample set; based on the comprehensive feature vector, a voiceprint recognition model is adaptively trained by using a sample selection strategy, and after training is completed, the trained voiceprint recognition model is applied to voiceprint recognition, so that the advantages of deep learning in the voiceprint recognition field are fully played, the voiceprint recognition speed and efficiency are improved, and the voiceprint recognition efficiency is improved. And the misrecognition rate and the missed recognition rate of the system are also reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of voiceprint recognition, and more specifically, to a feature-based intelligent voiceprint recognition method and system. Background Art

[0002] Voiceprint recognition has attracted much attention as a contactless means of biological or object identification; however, existing voiceprint recognition methods still have some shortcomings in practical applications; first, the existing methods have a major bottleneck in obtaining high-quality annotated voiceprint data, which requires the collection of biological or object sounds in various environments, and manual annotation by professionals, which is a huge workload and costly; due to the complexity and variability of environmental scenes, it is often difficult to obtain enough voiceprint samples to cover various situations, which seriously restricts the generalization ability of voiceprint recognition models; secondly, the existing voiceprint preprocessing method has rigid parameter settings and cannot adapt to the characteristics of different speech signals. For example, in a noisy indoor environment, due to the interference of background noise, fixed denoising parameters may not be able to effectively eliminate noise; in an outdoor environment, due to the influence of natural noise such as wind, fixed denoising parameters may not be able to effectively eliminate noise. The detection parameters may not be able to accurately detect the key voice parts, which will lead to poor preprocessing effects and affect the subsequent voiceprint recognition accuracy; furthermore, the existing voiceprint feature extraction methods are limited to a single dimension and cannot fully characterize the essential characteristics of the speech signal; they cannot fully characterize the dynamic change process of the speech signal, which will inevitably lead to information loss and affect the performance of voiceprint recognition; in addition, in the process of voiceprint recognition model training, the existing methods usually use full data set training or random sampling to select training samples, lacking an effective sample selection strategy, and full data set training easily leads to overfitting of the model to certain specific users and poor generalization ability; although random sampling can alleviate the overfitting problem to a certain extent, it still cannot ensure the generalization performance of the model because it does not consider the representativeness and diversity between samples, and the training efficiency is low.

[0003] In view of this, the present invention proposes a feature-based intelligent voiceprint recognition method and system to solve the above problems. Summary of the invention

[0004] In order to overcome the above defects of the prior art and to achieve the above objectives, the present invention provides the following technical solutions: A feature-based intelligent voiceprint recognition method, comprising: S1, collecting a number of voiceprint data samples, and preprocessing each voiceprint data sample to obtain a comprehensive voice sample; S2. constructing a speech directed graph model based on the comprehensive speech samples, and identifying a core speech sample set from the speech directed graph model; S3, extract features from the core speech sample set to obtain A comprehensive feature vector; S4, based on the comprehensive feature vector and the core speech sample set, using the clustering algorithm to obtain the sample selection strategy; S5. Based on the comprehensive feature vector, the voiceprint recognition model is adaptively trained using the sample selection strategy. After the training is completed, the trained voiceprint recognition model is applied to voiceprint recognition.

[0005] Furthermore, the pre-processing method includes: Each voiceprint data sample is denoised to obtain a denoised voiceprint sample, the silent segments before and after the denoised voiceprint sample are removed to obtain a standard voiceprint sample, and the standard voiceprint sample is pre-emphasized to obtain a comprehensive speech sample.

[0006] Furthermore, the denoising process includes: Select a silent segment from the voiceprint data sample, divide the frequency band of the silent segment into M sub-bands, each sub-band contains Nm frequency bins; For sub-bands, calculate the sub-band power spectrum of each frame ; ; in, is the index of the frame, is the index of the frequency bin, is a time domain series, is the time index, is the time index of exponential decay, is an exponential decay factor, and ; is the length of the preset exponential decay window; For sub-bands, calculate the average of the sub-band power spectra of the previous frames as the initial noise power spectrum estimate ; and the initial noise power spectrum estimate Make updates; Ways to update include: Set a low quantile threshold and a high quantile threshold ; For each frame of the subband , calculate the power spectrum value of each frequency bin of the frame ;Will With the initial noise power spectrum estimate Make comparisons; like ; Then update the noise power spectrum estimate of the next frame: ; like ; Then update the noise power spectrum estimate of the next frame: ; otherwise ; Smoothing the updated noise power spectrum estimation value to obtain the noise power spectrum density estimation value of each frame of each subband; Define a Wiener filter, and calculate the optimal coefficients of the Wiener filter based on the noise power spectral density estimate and the preset adaptive expected signal-to-noise ratio ; ;in, is the adaptive expected signal-to-noise ratio, For the Frequency domain coefficients of the frame; For the The noise power spectral density estimate of the frame; ; in, is the preset minimum expected signal-to-noise ratio, is the preset maximum expected signal-to-noise ratio, For the The spectral entropy of the frame, is the maximum spectral entropy; The voiceprint data samples are divided into frames and Fourier transformed to obtain the frequency domain coefficients of each frame; the frequency domain coefficients of each frame are multiplied by the optimal coefficients of the Wiener filter to obtain the filtered frequency domain coefficients; The filtered frequency domain coefficients are subjected to inverse Fourier transform to obtain the enhanced time domain signal of each frame; the enhanced time domain signals of all frames are overlapped and added to reconstruct the complete enhanced speech signal, which is the denoised ripple sample.

[0007] Furthermore, the silent segment is selected in the following manner: Divide the voiceprint data samples into several time frames; for each time frame, calculate its short-term energy En= β × ∑ a =0 A-1 [ y ( a ) - μ ] 2 +(1 - β ) × Enp ,in, represents the time domain sample sequence of the time frame, is the index of the sampling point of the time domain sample sequence, is the length of the time frame, is the mean value of the amplitude of the time domain sample sequence of the time frame, is the forgetting factor, and ; is the short-term energy of the previous time frame; Define the initial adaptive energy threshold , where is the mean value of the short-time energy of the voiceprint data sample; for each time frame, update the adaptive energy threshold of the current time frame to ; ; in, is the smoothing factor, is the adaptive energy threshold of the previous time frame, is the long-term statistical threshold; traverse all time frames, if the short-term energy of any time frame is less than the corresponding adaptive energy threshold, it will be marked as a silent frame; find the longest segment of continuous silent frames, record its starting frame number and ending frame number; start from the starting frame number and end with the ending frame number, extract the voiceprint data sample corresponding to this segment, which is the silent segment.

[0008] Furthermore, the method of performing the pre-emphasis processing includes: The standard voiceprint sample is pre-emphasized and filtered to obtain a pre-emphasized voice signal, and the pre-emphasized voice signal is amplitude normalized to complete the pre-emphasis processing to obtain a comprehensive voice sample; The formula for performing pre-emphasis filtering is: ; in, is the pre-emphasized speech signal, is the index of the time point; is the regulating factor, is the low-frequency component of the standard voiceprint sample, is the high frequency component of the standard voiceprint sample, For time point The sampling value of the standard voiceprint sample at For time point The sampling value of the standard voiceprint sample at For time point The sampling value of the standard voiceprint sample at ; is the second-order coefficient; is the pre-emphasis intensity control parameter, is the wavelet basis function; The method of performing amplitude normalization includes: The decomposition layer number L1 is preset, and the pre-emphasized speech signal is decomposed by L1 layer wavelet packet to obtain 2 L1wavelet packet components; calculate the energy of each wavelet packet component, that is, the sum of the squares of all amplitudes, calculate the sum of the energies of all wavelet packet components E_total, calculate the normalized weight W(l') of the l'th wavelet packet component, that is, the ratio of the energy of the corresponding wavelet packet component to the sum of the energies E_total; for the l'th wavelet packet component, obtain its maximum amplitude max_wpl' and minimum amplitude min_wpl'; scale the maximum amplitude max_wpl' and the minimum amplitude min_wpl' to the range of [-W(l'), W(l')] to obtain the normalized l'th wavelet packet component, reconstruct all the normalized wavelet packet components into wavelet packets, and obtain the final normalized speech signal.

[0009] Furthermore, the method of constructing the speech directed graph model includes: Each comprehensive speech sample is divided into frames to obtain a series of speech frames; a feature vector is extracted from each speech frame, and the feature vector is a Mel frequency cepstral coefficient or a linear prediction coefficient; Calculate the similarity of the feature vectors between any two speech frames as the similarity between the speech frames, and construct a weighted undirected graph based on the similarity between the speech frames ,in is a set of speech frames, is a set of edges, and the speech frames are used as nodes of a weighted undirected graph. If the similarity between two speech frames is less than a preset similarity threshold, the two nodes are connected in the weighted undirected graph, and the weight of the edge is defined as the similarity between the two speech frames; The similarity calculation formula is: ; in, There are two speech frames and The distance measure between For speech frames The short-term energy For another speech frame Short-term energy; and are two adjustment parameters; Two speech frames and The similarity between is the direction zero-crossing measurement function; Directional zero-crossing metric function ; In the formula, There are two speech frames and The angle between the eigenvectors of Represents two speech frames and The absolute value of the difference in the zero-crossing rate of is the zero-crossing adjustment parameter; Weighted Undirected Graph Perform a transpose operation to obtain its transposed graph ,Will The direction of each edge in is reversed to obtain a speech directed graph model; The ways to perform the transpose operation include: Build a graph neural network model, use the graph attention network as the skeleton network of the graph neural network model, and define the input of the graph neural network model as a weighted undirected graph The topological structure and nodes of the weighted undirected graph are output as low-dimensional continuous embedding representations of the nodes. For each node, its MFCC feature vector is extracted as the initial node feature. And its initial node features are input into the graph neural network model, and the training objective of the graph neural network model is defined as node classification or link prediction; The parameters of the graph neural network model are trained through the back-propagation algorithm. The graph neural network model learns to obtain the low-dimensional continuous embedding representation of each node. Based on the obtained low-dimensional continuous embedding representation of all nodes, the adjacency matrix A_emb of the embedding space is constructed; the elements of the adjacency matrix A_emb are the Euclidean distances of the low-dimensional continuous embedding representations between the nodes; the adjacency matrix A_emb is transposed to obtain the transposed adjacency matrix A_emb_T, and the directed graph is reconstructed according to the transposed adjacency matrix A_emb_T, that is, the transposed graph .

[0010] Furthermore, the recognition method of the core voice sample set includes: All nodes in the speech directed graph model are randomly arranged in any area on the two-dimensional plane as the initial state of all nodes; For any two nodes and , calculate the repulsive force, attractive force and directional force of the edge between the two; For each node , the repulsive force, attractive force and directional force acting on it are vector synthesized to obtain the resultant force ; According to the magnitude and direction of the resultant force, calculate the node displacement; The displacement is calculated as: ; in, is the displacement scaling factor, For the joint efforts Direction, is the preset desired direction, , , and is the displacement adjustment parameter; For Node The velocity vector of All nodes are moved according to the calculated displacement to obtain a new node layout; the energy of the new node layout is calculated, and if the energy is less than the preset iteration threshold, the iteration is terminated; at this time, several nodes are gathered in the central area of ​​any area, and the speech frames corresponding to the nodes in the central area are the core speech sample set.

[0011] Further, for each speech frame in the core speech sample set, its short-time square root amplitude, short-time energy, short-time average amplitude and short-time zero-crossing rate are extracted as time domain features; For each speech frame in the core speech sample set, extract its spectrum entropy, spectrum flatness, spectrum median, spectrum effective bandwidth and spectrum peak frequency as frequency domain features; The extracted time domain features and frequency domain features are concatenated at the vector level in a fixed order to construct a comprehensive feature vector of the corresponding speech frame; that is, A comprehensive feature vector; The method of obtaining the sample selection strategy includes: Take each speech frame in the core speech sample set as a sample, and the comprehensive feature vector corresponding to the sample as a row to construct a The characteristic matrix of is the number of samples, is the dimension of the comprehensive feature vector; The preset window radius R is used for each sample. , in is the center and the radius is Within the spherical range, calculate the sample The sample density ; ;in, are other samples within the spherical range, is the kernel function, is the volume of the spherical range, is the regularization parameter, for The sample weights of is the bandwidth parameter; The samples with a sample density greater than the preset minimum sample density threshold Min are marked as core samples; for any two core samples, if there is a series of samples such that the distance between any two adjacent core samples is less than R, then the two core samples are density-reachable; Construct a density-reachable matrix , where the elements M1[J 1 ,J 1' ]=1 Indicates the core sample and core samples is density-accessible; traverse the density-accessible matrix , all density-reachable core samples are grouped into one cluster; For each cluster, calculate the ratio of the number of samples in it to the total number of samples, which is recorded as the strategy ratio ; According to the strategy ratio , randomly select the corresponding samples; the process of randomly selecting samples from each cluster is called the sample selection strategy.

[0012] Furthermore, the method of adaptively training the voiceprint recognition model using the sample selection strategy includes: The samples randomly selected from each cluster and their corresponding comprehensive feature vectors are used as training data sets, and the samples are annotated according to the voiceprint type as labels of the training data sets; Define the deep neural network model structure as the basic architecture of the voiceprint recognition model; and initialize the parameters of the voiceprint recognition model; The comprehensive feature vectors corresponding to the samples randomly selected from any cluster are taken as a batch to train the voiceprint recognition model, and the back propagation algorithm and the optimization algorithm are used to update the parameters of the voiceprint recognition model. The training loss function is defined, and the performance of the voiceprint recognition model on the batch is evaluated according to the training loss function to determine whether the performance of the voiceprint recognition model on the batch meets the expectations. If so, enter the next round of iteration and continue to train the voiceprint recognition model using the comprehensive feature vectors corresponding to the samples randomly selected from the next cluster. If not, adjust the parameters of the voiceprint recognition model according to the gradient of the value of the training loss function during the training process, and continue training; until the number of training rounds reaches the preset training maximum value, stop training, and obtain a trained voiceprint recognition model.

[0013] A feature-based intelligent voiceprint recognition system, which is used to implement the feature-based intelligent voiceprint recognition method, includes: a voiceprint collection and processing module, which is used to collect a number of voiceprint data samples, pre-process each voiceprint data sample, and obtain a comprehensive voice sample; The graph model construction and screening module builds a speech directed graph model based on comprehensive speech samples and identifies a core speech sample set from the speech directed graph model; The feature extraction module is used to extract features from the core speech sample set to obtain A comprehensive feature vector; The strategy planning module uses clustering algorithm to obtain sample selection strategy based on comprehensive feature vectors and core speech sample sets; The model training and delivery module uses a sample selection strategy to adaptively train the voiceprint recognition model based on the comprehensive feature vector. After the training is completed, the trained voiceprint recognition model is applied to voiceprint recognition. The modules are connected by wired and / or wireless means.

[0014] The technical effects and advantages of the feature-based intelligent voiceprint recognition method and system of the present invention are as follows: The present invention improves the accuracy and reliability of voiceprint recognition, especially performs well in complex noise environments, greatly enhances the system's adaptability to different types of sound signals, and enables it to flexibly respond to various voiceprint recognition scenarios; secondly, by extracting comprehensive time domain and frequency domain features, it can more accurately capture the essential characteristics of speech signals, thereby significantly improving the accuracy of voiceprint recognition; in addition, through optimized sample selection and feature extraction strategies, it can more effectively utilize key information in the data set and improve the learning efficiency of the model. Its adaptive training mechanism enables the model to be dynamically optimized according to different data characteristics, significantly improving the generalization ability of the system; it is suitable for a wider range of voice scenarios; in addition, it also has strong anti-interference ability and environmental adaptability, and can maintain stable recognition performance under various complex conditions; it not only improves the speed and efficiency of voiceprint recognition, but also reduces the system's misrecognition rate and missed recognition rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram of a feature-based intelligent voiceprint recognition method of the present invention; Figure 2 A schematic diagram of a feature-based intelligent voiceprint recognition system of the present invention. DETAILED DESCRIPTION

[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] Example 1 See also Figure 1 As shown, this embodiment provides a feature-based intelligent voiceprint recognition method, including: S1. Collect several voiceprint data samples, and pre-process each voiceprint data sample to obtain a comprehensive voice sample; S2. constructing a speech directed graph model based on the comprehensive speech samples, and identifying a core speech sample set from the speech directed graph model; S3, extract features from the core speech sample set to obtain A comprehensive feature vector; S4, based on the comprehensive feature vector and the core speech sample set, using the clustering algorithm to obtain the sample selection strategy; S5. Based on the comprehensive feature vector, the voiceprint recognition model is adaptively trained using the sample selection strategy. After the training is completed, the trained voiceprint recognition model is applied to voiceprint recognition.

[0018] The methods for collecting voiceprint data samples include: Use on-site recording equipment to record and collect target sounds in actual environments, such as recording human voices in different occasions. It should be noted that during the collection process, the authenticity and diversity of voice samples should be ensured as much as possible to improve the robustness and accuracy of the voiceprint recognition model.

[0019] The methods for preprocessing each voiceprint data sample include: Each voiceprint data sample is denoised to obtain a denoised voiceprint sample. Endpoint detection is performed based on the denoised voiceprint sample, that is, the silent segments (soundless areas) before and after the denoised voiceprint sample are removed to obtain a standard voiceprint sample. The standard voiceprint sample is pre-emphasized to obtain a comprehensive speech sample.

[0020] The methods for denoising include: A silent segment is selected from the voiceprint data sample, and the frequency band of the silent segment is divided into M sub-bands, each sub-band contains Nm frequency bins; sub-bands, calculate the sub-band power spectrum of each frame ; ; in, is the index of the frame, is the index of the frequency bin, is a time domain sequence (the original voiceprint data sample is a continuous analog signal, which needs to be digitized through sampling to obtain a discrete digital sequence, which is the time domain sequence). is the time index, is the time index of exponential decay, Indicates that the time domain sequence is time-shifted. is an exponential decay factor, and ; The length of the preset exponential decay window.

[0021] For sub-bands, calculate the average of the sub-band power spectra of the previous frames as the initial noise power spectrum estimate ; and the initial noise power spectrum estimate to update.

[0022] Specifically, set a low quantile threshold and a high quantile threshold ,like , corresponding to the 20% low quantile; , corresponding to the 80% high quantile.

[0023] For each frame of the subband , calculate the power spectrum value of each frequency bin of the frame ;Will With the initial noise power spectrum estimate Make a comparison.

[0024] like , then update the noise power spectrum estimate of the next frame ; like , then update the noise power spectrum estimate of the next frame ;otherwise ; The updated noise power spectrum estimate is smoothed, such as by using a median filter or an exponential average, to obtain a noise power spectrum density estimate for each frame of each subband.

[0025] Define a Wiener filter, and calculate the optimal coefficients of the Wiener filter based on the noise power spectral density estimate and the preset adaptive expected signal-to-noise ratio ; ;in, is the adaptive expected signal-to-noise ratio, For the Frequency domain coefficients of the frame; For the Noise power spectral density estimate for the frame.

[0026] ; in, is the preset minimum expected signal-to-noise ratio, is the preset maximum expected signal-to-noise ratio, For the The spectral entropy of the frame, is the maximum spectral entropy.

[0027] The voiceprint data samples are divided into frames and Fourier transformed to obtain the frequency domain coefficients of each frame; the frequency domain coefficients of each frame are multiplied by the optimal coefficients of the Wiener filter to obtain the filtered frequency domain coefficients.

[0028] The filtered frequency domain coefficients are subjected to inverse Fourier transform to obtain the enhanced time domain signal of each frame; the enhanced time domain signals of all frames are overlapped and added to reconstruct the complete enhanced speech signal, which is the denoised ripple sample.

[0029] Furthermore, the selection method of the silent segment includes: The voiceprint data samples are divided into several time frames. Usually, the frame length of the time frame is 20-30ms, and the frame shift is about 10ms. For each time frame, its short-time energy is calculated. .

[0030] En= β × ∑ a =0 A-1 [ y ( a ) - μ ] 2 +(1 - β ) × Enp ,in, represents the time domain sample sequence of the time frame, is the index of the sampling point of the time domain sample sequence, is the length of the time frame, is the mean value (mean value of amplitude) of the time domain sample sequence of the time frame, is the forgetting factor, and ; Used to control the weight of the current frame energy and the previous frame energy; is the short-term energy of the previous time frame.

[0031] Define the initial adaptive energy threshold , where is the mean of the short-term energy of the voiceprint data samples.

[0032] For each time frame, update the adaptive energy threshold of the current time frame to ; ; in, is the smoothing factor, used to adjust the weight, is the adaptive energy threshold of the previous time frame, It is the long-term statistical threshold, which means the energy threshold obtained based on statistics of a longer time window.

[0033] Traverse all time frames. If the short-term energy of any time frame is less than the corresponding adaptive energy threshold, mark it as a silent frame. Find the longest segment of continuous silent frames and record its starting frame number and ending frame number. From the starting frame number to the ending frame number, extract the voiceprint data sample corresponding to this segment, which is the silent segment.

[0034] Further, the method of performing pre-emphasis processing includes: The standard voiceprint sample is pre-emphasized and filtered to obtain a pre-emphasized voice signal, and the pre-emphasized voice signal is amplitude normalized to complete the pre-emphasis processing and obtain a comprehensive voice sample.

[0035] The formula for pre-emphasis filtering is: ; in, is the pre-emphasized speech signal (sequence), is the index of the time point; is an adjustment factor used to balance the ratio of low-frequency and high-frequency components. is the low-frequency component of the standard voiceprint sample, is the high frequency component of the standard voiceprint sample, For time point The sampling value of the standard voiceprint sample at For time point The sampling value of the standard voiceprint sample at For time point The sampling value of the standard voiceprint sample at ; is the second-order coefficient, usually -0.73; is the pre-emphasis intensity control parameter, As the wavelet basis function, a wavelet basis function with good high-pass characteristics can be selected.

[0036] The decomposition layer number L1 is preset, and the pre-emphasized speech signal is decomposed by L1 layer wavelet packet (selecting a suitable wavelet basis function, such as Daubechies wavelet, etc.), and 2 L1wavelet packet components; calculate the energy of each wavelet packet component, that is, the sum of the squares of all amplitudes, calculate the sum of the energies of all wavelet packet components E_total, calculate the normalized weight W(l') of the l'th wavelet packet component, that is, the ratio of the energy of the corresponding wavelet packet component to the sum of the energies E_total; for the l'th wavelet packet component, obtain its maximum amplitude max_wpl' and minimum amplitude min_wpl'; scale the maximum amplitude max_wpl' and the minimum amplitude min_wpl' to the range of [-W(l'), W(l')] to obtain the normalized l'th wavelet packet component, reconstruct all the normalized wavelet packet components into wavelet packets, and obtain the final normalized speech signal.

[0037] The pre-emphasis processing process becomes more intelligent and adaptive, and can dynamically adjust parameters according to the characteristics of different speech signals, thereby obtaining a better pre-emphasis effect.

[0038] Furthermore, the methods of constructing a speech directed graph model include: Each comprehensive speech sample is divided into frames to obtain a series of speech frames; a feature vector is extracted from each speech frame, and the feature vector is a Mel-frequency cepstral coefficient (MFCC) or a linear prediction coefficient (LPC).

[0039] Calculate the similarity of the feature vectors between any two speech frames as the similarity between the speech frames, and construct a weighted undirected graph based on the similarity between the speech frames ,in is a set of speech frames, is a set of edges, and the speech frames are taken as nodes of a weighted undirected graph. If the similarity between two speech frames is less than a preset similarity threshold, the two nodes are connected in the weighted undirected graph, and the weight of the edge is defined as the similarity between the two speech frames.

[0040] The similarity calculation formula is: ; in, There are two speech frames and The distance measure between (choose Euclidean distance or Manhattan distance); For speech frames The short-term energy For another speech frame Short-term energy; and are two adjustment parameters; used to control the influence of each item on similarity; Two speech frames and The similarity between is the direction zero-crossing measurement function; Directional zero-crossing metric function ; In the formula, There are two speech frames and The angle between the eigenvectors of Represents two speech frames and The absolute value of the difference in the zero-crossing rate of It is the zero-crossing adjustment parameter; it is used to control the weight of the zero-crossing rate term.

[0041] Weighted Undirected Graph Perform a transpose operation to obtain its transposed graph ,Will The direction of each edge in is reversed to obtain a speech directed graph model.

[0042] The ways to perform the transpose operation include: Build a graph neural network model, use the graph attention network (GAT) as the skeleton network of the graph neural network model, and define the input of the graph neural network model as a weighted undirected graph The topological structure and nodes of the weighted undirected graph are: the output is a low-dimensional continuous embedding representation of the node; for each node (speech frame), its MFCC feature vector is extracted as the initial node feature; and its initial node features are input into the graph neural network model, and the training objective of the graph neural network model is defined as node classification or link prediction; specifically, the training objective of node classification is to predict which category the node belongs to based on the low-dimensional continuous embedding representation of the node, and a category label is assigned to each node in advance; the model maps the embedding representation of each node to the corresponding category label; by minimizing the loss function of node classification (such as cross entropy loss), the model learns the embedding representation that can well distinguish nodes of different categories.

[0043] The training goal is link prediction, that is, the edge (connection) between nodes is taken as the target to be predicted; the task of the model is to predict whether there is an edge (connection) between two nodes based on their embedded representations; specifically, if the similarity between two nodes is high, then there is likely to be an edge between them, conversely, if the similarity between two nodes is low, then there may not be an edge between them; during the training process, the model will learn to map the embedded representations of two nodes to a similarity score; by minimizing the loss function of link prediction (such as binary cross entropy loss), the model can learn an embedded representation that can well predict the connection relationship between nodes.

[0044] The parameters of the graph neural network model are trained through the back-propagation algorithm. The graph neural network model learns to obtain the low-dimensional continuous embedding representation of each node. Based on the obtained low-dimensional continuous embedding representation of all nodes, the adjacency matrix A_emb of the embedding space is constructed; the elements of the adjacency matrix A_emb are the Euclidean distances of the low-dimensional continuous embedding representations between the nodes; the adjacency matrix A_emb is transposed to obtain the transposed adjacency matrix A_emb_T, and the directed graph is reconstructed according to the transposed adjacency matrix A_emb_T, that is, the transposed graph .

[0045] Furthermore, the recognition method of the core speech sample set includes: All nodes (speech frames) in the speech directed graph model are randomly arranged in any area on the two-dimensional plane as the initial state of all nodes; for any two nodes and , calculate the repulsive force, attractive force and directional force of the edge between the two; The formula for calculating the repulsive force is: ; in, For two nodes and The repulsive force between is the repulsion coefficient, which is used to adjust the magnitude of the repulsion. is the repulsive force constant, controlling the attenuation rate of the repulsive force, For two nodes and The Euclidean distance between is the repulsion adjustment parameter, which is used to control the influence of similarity on repulsion; For two nodes and The similarity between is calculated in the same way).

[0046] The formula for calculating gravity is: ; in, For two nodes and The attraction between is the gravitational constant, controlling the decay rate of gravity, For Node and The weight of the edge between them reflects the degree of connection between the two nodes; For Node The zero-crossing rate of the corresponding speech frame, For Node The zero-crossing rate of the corresponding speech frame, Gravity adjustment parameter; used to control the influence of similarity on gravity.

[0047] The calculation formula for the directional force of the edge is: ; in, For Node and Directional force between edges; For Node and The angle between (substitute the angle of the eigenvector into the calculation); is the hyperbolic tangent function, and its range is (-1, 1); is the similarity influencing parameter, It is the distance influence parameter, which is used to control the influence of similarity and distance on the directional force. It is the directional force coefficient; it is used to adjust the magnitude of the directional force.

[0048] For each node , the repulsive force, attractive force and directional force acting on it are vector synthesized to obtain the resultant force ; According to the magnitude and direction of the resultant force, calculate the node displacement (length and direction of movement).

[0049] The displacement is calculated as: ; in, is the displacement scaling factor, For the joint efforts Direction, is the preset desired direction (such as horizontal direction), , , and It is the displacement adjustment parameter, which controls the influence of several information on the displacement; For Node The velocity vector can be expressed as A weighted sum of the displacement vectors in the first few iterations is calculated.

[0050] Move all nodes according to the calculated displacement to obtain a new node layout; calculate the energy of the new node layout (the sum of the gravitational force and the gravitational force between nodes). If the energy is less than the preset iteration threshold, the iteration is terminated; at this time, several nodes are gathered in the central area of ​​any area (pre-divided) and are closely connected to each other. The speech frames corresponding to these nodes are the core speech sample set.

[0051] First, some typical speech samples are tested and the curve of node layout energy changing with the number of iterations is plotted; when the energy curve tends to be flat and the rate of change is lower than a certain threshold, the threshold can be used as the iteration threshold.

[0052] Further, the feature extraction method includes: For each speech frame in the core speech sample set, its short-time square root amplitude, short-time energy, short-time average amplitude and short-time zero-crossing rate are extracted as time domain features.

[0053] For each speech frame in the core speech sample set, its spectral entropy, spectral flatness, spectral median, spectral effective bandwidth and spectral peak frequency are extracted as frequency domain features; the extracted time domain features and frequency domain features are vector-level spliced ​​in a fixed order to construct a comprehensive feature vector of the corresponding speech frame; that is, A comprehensive feature vector.

[0054] Furthermore, the sample selection strategy is obtained by: Take each speech frame in the core speech sample set as a sample, and the comprehensive feature vector corresponding to the sample as a row to construct a The characteristic matrix of is the number of samples, is the dimension of the comprehensive feature vector.

[0055] Preset window radius , for each sample , in is the center and the radius is Within the spherical range, calculate the sample The sample density ; ;in, are other samples within the spherical range, is a kernel function (such as a Gaussian kernel), is the volume of the spherical range, is a regularization parameter used to avoid the situation where the denominator is 0. for The sample weight can be a confidence score, a manually annotated confidence, or a model-predicted confidence probability. A higher weight is given to samples with higher quality, thereby improving the accuracy of density estimation. is the bandwidth parameter, which controls the sliding window size of the kernel function.

[0056] The samples whose sample density is greater than the preset minimum sample density threshold Min are marked as core samples; for any two core samples, if there is a density chain connecting them, that is, there is a series of samples such that the distance between any two adjacent samples is less than R, then the two core samples are density reachable, that is, there is a density chain.

[0057] Construct a density-reachable matrix , where the elements M1[J 1 ,J 1' ]=1 Indicates the core sample and core samples is density-accessible; traverse the density-accessible matrix , all density-reachable core samples are grouped into one cluster.

[0058] For each cluster, calculate the ratio of the number of samples in it to the total number of samples, which is recorded as the strategy ratio ; According to the strategy ratio , randomly select the corresponding samples; the process of randomly selecting samples from each cluster is called the sample selection strategy.

[0059] Specifically, the methods of adaptively training the voiceprint recognition model using the sample selection strategy include: The samples randomly selected from each cluster and their corresponding comprehensive feature vectors are used as the training data set, and the samples are annotated according to the voiceprint type (such as wind sound, breaking glass sound, explosion sound, etc.) as the labels of the training data set.

[0060] Define the deep neural network model structure, such as convolutional neural network (CNN), recurrent neural network (RNN) or a combination of the two, as the basic architecture of the voiceprint recognition model; and initialize the parameters of the voiceprint recognition model, including the number of network layers, number of nodes, and activation function.

[0061] The comprehensive feature vectors corresponding to the samples randomly selected from any cluster are taken as a batch to train the voiceprint recognition model, and the back propagation algorithm and optimization algorithm (such as SGD, Adam, etc.) are used to update the parameters of the voiceprint recognition model, and the training loss function (such as cross entropy) is defined. The performance of the voiceprint recognition model on the batch is evaluated according to the training loss function. The performance is reflected by the accuracy, that is, (the number of correctly predicted samples) / (the total number of samples), to determine whether the performance of the voiceprint recognition model on the batch meets expectations, that is, the accuracy is greater than the preset accuracy threshold; if it meets the requirements, enter the next round of iteration, and continue to train the voiceprint recognition model using the comprehensive feature vectors corresponding to the samples randomly selected from the next cluster; if it does not meet the requirements, adjust the parameters of the voiceprint recognition model according to the gradient of the value of the training loss function during the training process, and continue training; until the number of training rounds reaches the preset training maximum value, stop training, and obtain the trained voiceprint recognition model; apply the trained voiceprint recognition model to the actual voiceprint recognition task; input the speech sample to be recognized, and the voiceprint recognition model will output the corresponding voiceprint type prediction result.

[0062] This embodiment improves the accuracy and reliability of voiceprint recognition, especially performs well in complex noise environments, greatly enhances the system's adaptability to different types of sound signals, and enables it to flexibly respond to various voiceprint recognition scenarios; secondly, by extracting comprehensive time domain and frequency domain features, it can more accurately capture the essential characteristics of speech signals, thereby significantly improving the accuracy of voiceprint recognition; in addition, through optimized sample selection and feature extraction strategies, it can more effectively utilize key information in the data set and improve the learning efficiency of the model. Its adaptive training mechanism enables the model to be dynamically optimized according to different data characteristics, significantly improving the generalization ability of the system; it is suitable for a wider range of voice scenarios; in addition, it also has strong anti-interference ability and environmental adaptability, and can maintain stable recognition performance under various complex conditions; it not only improves the speed and efficiency of voiceprint recognition, but also reduces the system's false recognition rate and missed recognition rate.

[0063] Example 2 See also Figure 2 As shown, the part not described in detail in this embodiment can be found in the description of Embodiment 1, which provides a feature-based intelligent voiceprint recognition system, including: The voiceprint collection and processing module is used to collect a number of voiceprint data samples, pre-process each voiceprint data sample, and obtain a comprehensive voice sample; The graph model construction and screening module builds a speech directed graph model based on comprehensive speech samples and identifies a core speech sample set from the speech directed graph model; The feature extraction module is used to extract features from the core speech sample set to obtain A comprehensive feature vector; The strategy planning module uses clustering algorithm to obtain sample selection strategy based on comprehensive feature vectors and core speech sample sets; The model training and delivery module uses sample selection strategies to adaptively train the voiceprint recognition model based on the comprehensive feature vector. After the training is completed, the trained voiceprint recognition model is applied to voiceprint recognition. The modules are connected by wired and / or wireless means to achieve data transmission between modules.

[0064] Example 3 This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the operation mode of the feature-based intelligent voiceprint recognition method provided above is implemented.

[0065] Since the electronic device introduced in this embodiment is an electronic device used to implement a feature-based intelligent voiceprint recognition method in the embodiment of this application, based on the feature-based intelligent voiceprint recognition method introduced in the embodiment of this application, a person skilled in the art can understand the specific implementation of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application is not described in detail here. As long as a person skilled in the art implements an electronic device used in a feature-based intelligent voiceprint recognition method in the embodiment of this application, it belongs to the scope of protection of this application.

[0066] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0067] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A feature-based intelligent voiceprint recognition method, characterized in that: include: S1. Collect several voiceprint data samples, and pre-process each voiceprint data sample to obtain a comprehensive voice sample; S2. constructing a speech directed graph model based on the comprehensive speech samples, and identifying a core speech sample set from the speech directed graph model; S3, extract features from the core speech sample set to obtain A comprehensive feature vector; S4, based on the comprehensive feature vector and the core speech sample set, using the clustering algorithm to obtain the sample selection strategy; S5. Based on the comprehensive feature vector, the voiceprint recognition model is adaptively trained using the sample selection strategy. After the training is completed, the trained voiceprint recognition model is applied to voiceprint recognition.

2. The feature-based intelligent voiceprint recognition method according to claim 1, characterized in that: The method of pre-processing comprises: Each voiceprint data sample is denoised to obtain a denoised voiceprint sample, the silent segments before and after the denoised voiceprint sample are removed to obtain a standard voiceprint sample, and the standard voiceprint sample is pre-emphasized to obtain a comprehensive speech sample.

3. The feature-based intelligent voiceprint recognition method according to claim 2, characterized in that: The method of performing denoising processing includes: Select a silent segment from the voiceprint data sample, divide the frequency band of the silent segment into M sub-bands, each sub-band contains Nm frequency bins; For sub-bands, calculate the sub-band power spectrum of each frame ; ； in, is the index of the frame, is the index of the frequency bin, is a time domain series, is the time index, is the time index of exponential decay, is an exponential decay factor, and ; is the length of the preset exponential decay window; For sub-bands, calculate the average of the sub-band power spectra of the previous several frames as the initial noise power spectrum estimate ; and the initial noise power spectrum estimate Make updates; Ways to update include: Set a low quantile threshold and a high quantile threshold ; For each frame of the subband , calculate the power spectrum value of each frequency bin of the frame ;Will With the initial noise power spectrum estimate Make comparisons; like ; Then update the noise power spectrum estimate of the next frame: ； like ; Then update the noise power spectrum estimate of the next frame: ； otherwise ; Smoothing the updated noise power spectrum estimation value to obtain the noise power spectrum density estimation value of each frame of each subband; Define a Wiener filter, and calculate the optimal coefficients of the Wiener filter based on the noise power spectral density estimate and the preset adaptive expected signal-to-noise ratio ; ;in, is the adaptive expected signal-to-noise ratio, For the Frequency domain coefficients of the frame; For the The noise power spectral density estimate of the frame; ； in, is the preset minimum expected signal-to-noise ratio, is the preset maximum expected signal-to-noise ratio, For the The spectral entropy of the frame, is the maximum spectral entropy; The voiceprint data samples are divided into frames and Fourier transformed to obtain the frequency domain coefficients of each frame; the frequency domain coefficients of each frame are multiplied by the optimal coefficients of the Wiener filter to obtain the filtered frequency domain coefficients; The filtered frequency domain coefficients are subjected to inverse Fourier transform to obtain the enhanced time domain signal of each frame; the enhanced time domain signals of all frames are overlapped and added to reconstruct the complete enhanced speech signal, which is the denoised ripple sample.

4. The feature-based intelligent voiceprint recognition method according to claim 3 is characterized in that: The method for selecting the silent segment includes: Divide the voiceprint data samples into several time frames; for each time frame, calculate its short-term energy ,in, represents the time domain sample sequence of the time frame, is the index of the sampling point of the time domain sample sequence, is the length of the time frame, is the mean value of the amplitude of the time domain sample sequence of the time frame, is the forgetting factor, and ; is the short-term energy of the previous time frame; Define the initial adaptive energy threshold , where is the mean value of the short-time energy of the voiceprint data sample; for each time frame, update the adaptive energy threshold of the current time frame to ; ； in, is the smoothing factor, is the adaptive energy threshold of the previous time frame, is the long-term statistical threshold; traverse all time frames, if the short-term energy of any time frame is less than the corresponding adaptive energy threshold, it will be marked as a silent frame; find the longest segment of continuous silent frames, record its starting frame number and ending frame number; start from the starting frame number and end with the ending frame number, extract the voiceprint data sample corresponding to this segment, which is the silent segment.

5. The feature-based intelligent voiceprint recognition method according to claim 4, characterized in that: The method of performing the pre-emphasis processing includes: The standard voiceprint sample is pre-emphasized and filtered to obtain a pre-emphasized voice signal, and the pre-emphasized voice signal is amplitude normalized to complete the pre-emphasis processing to obtain a comprehensive voice sample; The formula for performing pre-emphasis filtering is: ； in, is the pre-emphasized speech signal, is the index of the time point; is the regulating factor, is the low-frequency component of the standard voiceprint sample, is the high frequency component of the standard voiceprint sample, For time point The sampling value of the standard voiceprint sample at For time point The sampling value of the standard voiceprint sample at For time point The sampling value of the standard voiceprint sample at ; is the second-order coefficient; is the pre-emphasis intensity control parameter, is the wavelet basis function; The method of performing amplitude normalization includes: The decomposition layer number L1 is preset, and the pre-emphasized speech signal is decomposed by L1 layer wavelet packet to obtain 2 L1 wavelet packet components; calculate the energy of each wavelet packet component, that is, the sum of the squares of all amplitudes, calculate the sum of the energies of all wavelet packet components E_total, calculate the normalized weight W(l') of the l'th wavelet packet component, that is, the ratio of the energy of the corresponding wavelet packet component to the sum of the energies E_total; for the l'th wavelet packet component, obtain its maximum amplitude max_wpl' and minimum amplitude min_wpl'; scale the maximum amplitude max_wpl' and the minimum amplitude min_wpl' to the range of [-W(l'), W(l')] to obtain the normalized l'th wavelet packet component, reconstruct all the normalized wavelet packet components into wavelet packets, and obtain the final normalized speech signal.

6. The feature-based intelligent voiceprint recognition method according to claim 5, characterized in that: The method of constructing the speech directed graph model includes: Each comprehensive speech sample is divided into frames to obtain a series of speech frames; a feature vector is extracted from each speech frame, and the feature vector is a Mel frequency cepstral coefficient or a linear prediction coefficient; Calculate the similarity of the feature vectors between any two speech frames as the similarity between the speech frames, and construct a weighted undirected graph based on the similarity between the speech frames ,in is a set of speech frames, is a set of edges, and the speech frames are taken as nodes of a weighted undirected graph. If the similarity between two speech frames is less than a preset similarity threshold, the two nodes are connected in the weighted undirected graph, and the weight of the edge is defined as the similarity between the two speech frames; The similarity calculation formula is: ； in, There are two speech frames and The distance measure between For speech frames The short-term energy For another speech frame Short-term energy; and are two adjustment parameters; Two speech frames and The similarity between is the direction zero-crossing measurement function; Directional zero-crossing metric function ; In the formula, There are two speech frames and The angle between the eigenvectors of Represents two speech frames and The absolute value of the difference in the zero-crossing rate of is the zero-crossing adjustment parameter; Weighted Undirected Graph Perform a transpose operation to obtain its transposed graph ,Will The direction of each edge in is reversed to obtain a speech directed graph model; The ways to perform the transpose operation include: Build a graph neural network model, use the graph attention network as the skeleton network of the graph neural network model, and define the input of the graph neural network model as a weighted undirected graph The topological structure and nodes of the weighted undirected graph are output as low-dimensional continuous embedding representations of the nodes. For each node, its MFCC feature vector is extracted as the initial node feature. And its initial node features are input into the graph neural network model, and the training objective of the graph neural network model is defined as node classification or link prediction; The parameters of the graph neural network model are trained through the back-propagation algorithm. The graph neural network model learns to obtain the low-dimensional continuous embedding representation of each node. Based on the obtained low-dimensional continuous embedding representation of all nodes, the adjacency matrix A_emb of the embedding space is constructed; the elements of the adjacency matrix A_emb are the Euclidean distances of the low-dimensional continuous embedding representations between the nodes; the adjacency matrix A_emb is transposed to obtain the transposed adjacency matrix A_emb_T, and the directed graph is reconstructed according to the transposed adjacency matrix A_emb_T, that is, the transposed graph .

7. The feature-based intelligent voiceprint recognition method according to claim 6, characterized in that: The recognition method of the core voice sample set includes: All nodes in the speech directed graph model are randomly arranged in any area on the two-dimensional plane as the initial state of all nodes; For any two nodes and , calculate the repulsive force, attractive force and directional force of the edge between the two; For each node , the repulsive force, attractive force and directional force acting on it are vector synthesized to obtain the resultant force ; According to the magnitude and direction of the resultant force, calculate the node displacement; The displacement is calculated as: ； in, is the displacement scaling factor, For the joint efforts Direction, is the preset desired direction, , , and is the displacement adjustment parameter; For Node The velocity vector of All nodes are moved according to the calculated displacement to obtain a new node layout; the energy of the new node layout is calculated, and if the energy is less than the preset iteration threshold, the iteration is terminated; at this time, several nodes are gathered in the central area of ​​any area, and the speech frames corresponding to the nodes in the central area are the core speech sample set.

8. The feature-based intelligent voiceprint recognition method according to claim 7, characterized in that: For each speech frame in the core speech sample set, extract its short-time square root amplitude, short-time energy, short-time average amplitude and short-time zero-crossing rate as time domain features; For each speech frame in the core speech sample set, extract its spectrum entropy, spectrum flatness, spectrum median, spectrum effective bandwidth and spectrum peak frequency as frequency domain features; The extracted time domain features and frequency domain features are concatenated at the vector level in a fixed order to construct a comprehensive feature vector of the corresponding speech frame; that is, A comprehensive feature vector; The method of obtaining the sample selection strategy includes: Take each speech frame in the core speech sample set as a sample, and the comprehensive feature vector corresponding to the sample as a row to construct a The characteristic matrix of is the number of samples, is the dimension of the comprehensive feature vector; The preset window radius R is used for each sample. , in is the center and the radius is Within the spherical range, calculate the sample The sample density ; ;in, are other samples within the spherical range, is the kernel function, is the volume of the spherical range, is the regularization parameter, for The sample weights of is the bandwidth parameter; The samples with a sample density greater than the preset minimum sample density threshold Min are marked as core samples; for any two core samples, if there is a series of samples such that the distance between any two adjacent core samples is less than R, then the two core samples are density-reachable; Construct a density-reachable matrix , where the elements Indicates the core sample and core samples is density-accessible; traverse the density-accessible matrix , all density-reachable core samples are grouped into one cluster; For each cluster, calculate the ratio of the number of samples in it to the total number of samples, which is recorded as the strategy ratio ; According to the strategy ratio , randomly select the corresponding samples; the process of randomly selecting samples from each cluster is called the sample selection strategy.

9. The feature-based intelligent voiceprint recognition method according to claim 8, characterized in that: The method of adaptively training the voiceprint recognition model using the sample selection strategy includes: The samples randomly selected from each cluster and their corresponding comprehensive feature vectors are used as training data sets, and the samples are annotated according to the voiceprint type as labels of the training data sets; Define the deep neural network model structure as the basic architecture of the voiceprint recognition model; and initialize the parameters of the voiceprint recognition model; The comprehensive feature vectors corresponding to the samples randomly selected from any cluster are taken as a batch to train the voiceprint recognition model, and the back propagation algorithm and the optimization algorithm are used to update the parameters of the voiceprint recognition model. The training loss function is defined, and the performance of the voiceprint recognition model on the batch is evaluated according to the training loss function to determine whether the performance of the voiceprint recognition model on the batch meets the expectations. If so, enter the next round of iteration and continue to train the voiceprint recognition model using the comprehensive feature vectors corresponding to the samples randomly selected from the next cluster. If not, adjust the parameters of the voiceprint recognition model according to the gradient of the value of the training loss function during the training process, and continue training; until the number of training rounds reaches the preset training maximum value, stop training, and obtain a trained voiceprint recognition model.

10. A feature-based intelligent voiceprint recognition system, used to implement the feature-based intelligent voiceprint recognition method according to any one of claims 1 to 9, characterized in that: include: The voiceprint collection and processing module is used to collect a number of voiceprint data samples, pre-process each voiceprint data sample, and obtain a comprehensive voice sample; The graph model construction and screening module builds a speech directed graph model based on comprehensive speech samples and identifies a core speech sample set from the speech directed graph model; The feature extraction module is used to extract features from the core speech sample set to obtain A comprehensive feature vector; The strategy planning module uses clustering algorithm to obtain sample selection strategy based on comprehensive feature vectors and core speech sample sets; The model training and delivery module uses a sample selection strategy to adaptively train the voiceprint recognition model based on the comprehensive feature vector. After the training is completed, the trained voiceprint recognition model is applied to voiceprint recognition. The modules are connected by wired and / or wireless means.

Images