Language model-based dialect recognition method

Abstract

The invention discloses a dialect recognition method based on a language model, which can automatically extract key local spectrum features and a time sequence dependency relationship from a dialect voice signal, and avoids manual dependence on feature extraction in a traditional method, thereby remarkably improving the recognition precision of dialects. Through combination of a convolutional neural network (CNN) and a bidirectional gating loop unit (BiGRU), local details and long-time dependence features in voice signals can be deeply mined, and accurate capture of complex voice features by the model is ensured. In addition, through multi-level feature fusion, features of different levels are effectively combined, and the recognition capability of the model on the pronunciation difference between dialects is further improved. In the face of diversified dialect data, the method shows excellent robustness and adaptability, and can effectively cope with complex factors such as pronunciation differences and voice changes, thereby realizing higher accuracy and stability in dialect classification tasks.

Description

Technical Field

[0001] The invention belongs to the field of dialect recognition, and in particular, relates to a dialect recognition method based on a language model. Background Art

[0002] The task of dialect recognition faces many challenges, mainly due to the nonlinear and non-stationary nature of speech signals and the pronunciation differences between different dialects. The spectrum of speech signals varies over time, and features such as pronunciation, pitch, and speaking rate often vary significantly among dialects, making it difficult for traditional dialect recognition methods to effectively handle these complexities. Traditional methods based on the Gaussian mixture universal background model (GMM-UBM), although able to model to a certain extent, usually rely on handcrafted features such as Mel-frequency cepstral coefficients (MFCCs), which have limitations in processing non-stationary signals, especially when the pronunciation differences between different dialects are large, and the recognition accuracy is often low.

[0003] In recent years, deep learning methods, especially convolutional neural networks (CNN) and bidirectional gated recurrent units (BiGRU) have made significant progress in dialect recognition. CNN can automatically learn local features of speech signals, while BiGRU can effectively process time series information, making up for the shortcomings of traditional methods in time series modeling. However, these methods still have difficulties in processing non-stationary signal features, especially when faced with the diversity and complexity of different dialects. Although deep learning models are powerful, they may still be limited by the quality of data and the model's ability to express complex signal features.

[0004] In view of this, the present invention is proposed. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the deficiencies of the prior art and provide a dialect recognition method based on a language model, thereby solving the problems raised in the above-mentioned background technology.

[0006] In order to solve the above technical problems, the basic concept of the technical solution adopted by the present invention is:

[0007] A dialect recognition method based on a language model comprises the following steps:

[0008] S1: Preprocess the input dialect speech signal to obtain time-frequency features;

[0009] S2: inputting the time-frequency features into an empirical mode decomposition (EMD) module, decomposing the dialect speech signal into multiple empirical mode components (IMFs) and extracting the logarithmic energy features of each IMF;

[0010] S3: Fusing the logarithmic energy feature extracted in step S2 with the time-frequency feature extracted in step S1 to obtain a multi-dimensional feature vector for subsequent deep learning model training;

[0011] S4: input the fused features obtained in step S3 into a convolutional neural network (CNN), and extract the local spectrum features of the dialect signal by using a convolution operation;

[0012] S5: Input the local spectrum features extracted by CNN in step S4 into a bidirectional gated recurrent unit (BiGRU) module to extract the temporal features of the dialect speech;

[0013] S6: Input the temporal features extracted by BiGRU in step S5 into the Gaussian mixture general background model to classify and identify the dialect;

[0014] S7: Determine the dialect category based on the log-likelihood probability difference between the improved GMM-UBM model and the background model (UBM) in step S6, and output the final recognition result.

[0015] Optionally, the time-frequency feature is extracted by Mel-frequency cepstral coefficient method to obtain the feature vector F of each frame of speech t , where t represents the time frame index, And d is the dimension of the feature vector of each frame.

[0016] Optionally, the time-frequency features are input into an empirical mode decomposition (EMD) module, the dialect speech signal is decomposed into a plurality of empirical mode components (IMFs), and the steps of extracting the logarithmic energy features of each IMF are as follows:

[0017] S201: Determine the local maximum and local minimum of the original speech signal X(t), fit the upper and lower envelopes by cubic spline interpolation, obtain the upper envelope U(t) and the lower envelope L(t), and then, by formula Calculate the mean curve m1(t) of the envelope line, and then calculate the empirical mode component h1(t) according to the obtained mean curve m1(t) by the formula h1(t)=X(t)-m1(t);

[0018] S202: If h1(t) does not meet the preset threshold standard, repeat step S201 to obtain a new empirical mode component until the component meets the standard;

[0019] S203: Extracting logarithmic energy features E from all empirical mode components IMF (t), where the logarithmic energy feature is calculated as: Among them, h i (t) represents the i-th empirical mode component.

[0020] Optionally, the fusion feature obtained in step S3 is input into a convolutional neural network (CNN), and the steps of extracting the local spectrum feature of the dialect signal by using the convolution operation are as follows:

[0021] S401: Input the feature vector F obtained in step S3 and pass it to the convolution layer. The convolution operation extracts local features through the kernel function K to obtain the feature map F CNN , whose expression is: Among them, F i represents the i-th element in the input feature vector, K i is the convolution kernel, n is the size of the convolution kernel;

[0022] S402: Downsample the convolutional feature map through a pooling layer, usually using a maximum pooling operation, to obtain a feature map P CNN Then, the pooled feature map is flattened to obtain the vector P CNN-flat , so as to be input into the subsequent fully connected layer.

[0023] Optionally, the local spectrum features extracted by CNN in step S4 are input into a bidirectional gated recurrent unit (BiGRU) module to extract the temporal features of the dialect speech in the following steps:

[0024] S501: Input feature vector P CNN-flat The signal is passed to the bidirectional gated recurrent unit (BiGRU), which processes the signal forward and backward at the same time, obtains the front and back time sequence information of the dialect speech, and outputs the time sequence feature F BiGRU , the calculation process of BiGRU is: GRU forward =σ(W f F input +b f ), GRU backward =σ(W b F input +b b ) where σ represents the activation function, W f and W b are the weight matrices of the forward and backward GRUs, respectively, input is the input feature, b f and b b is the bias term.

[0025] S502: Combine the forward and backward GRU outputs obtained in step S501 to obtain the bidirectional temporal feature F BiGRU , which is used for subsequent classification decisions.

[0026] Optionally, the steps of inputting the time series features extracted by BiGRU in step S5 into the Gaussian mixture general background model to classify and identify the dialect are:

[0027] S601: Using a large-scale dialect dataset to train a universal background model (UBM) to obtain a Gaussian mixture model (GMM);

[0028] S602: Adaptively adjust the UBM using the Maximum A Posteriori Adaptive (MAP) algorithm to obtain a model λ suitable for the target dialect target ;

[0029] S603: By screening the training samples, the dialect model is optimized and its discrimination is enhanced, and finally a dialect model with stronger adaptability is obtained.

[0030] Optionally, dialect classification is performed based on the log-likelihood probability difference between the improved GMM-UBM model and the background model (UBM) in step S6, and its expression is: i (X) = logp(X|λ target )-logp(X|λ UBM ) where λ target is the target dialect model, λ UBM is the background model, and X is the input feature vector.

[0031] Calculate the log-likelihood Λ for each class i i (X), and determine the dialect category by maximizing the log-likelihood value. Finally, the log-likelihood probability difference is used to select the dialect category with the largest score and output the final recognition result.

[0032] Optionally, the steps of establishing a Gaussian mixture general background model are:

[0033] S6001: The time-frequency feature data {F1, F2, ..., F T} is input into the Gaussian mixture model for modeling. First, the mean vector μ and covariance matrix ∑ of the feature data are calculated.

[0034] S6002: Use the expectation maximization (EM) algorithm to train the GMM and estimate the parameters of the GMM by maximizing the likelihood function. The probability density function of the GMM is: Where λ = {π i , μ i ,∑ i} represents the parameters of GMM, π i is the weight of the ith Gaussian distribution, μ i and∑ i are the mean vector and covariance matrix of the i-th Gaussian distribution respectively.

[0035] S6003: Initialize GMM parameter π i , μ i ,∑ iUse the K-means clustering algorithm to perform preliminary clustering on the input features, and obtain the mean and covariance matrix of each cluster as the initial parameters of GMM;

[0036] S6004: Perform K-means clustering on the data set and calculate the mean of each cluster and the covariance matrix As initial GMM parameters;

[0037] S6005: Execute the EM algorithm to estimate parameters and calculate the responsibility γ through the expected step it , whose expression is: Among them, γ it Indicates the responsibility of the i-th Gaussian component to the t-th feature;

[0038] S6006: Update the parameters of GMM according to the degree of responsibility in step S6005, and update the weight π i , mean μ i , and the covariance matrix ∑ i , update the parameters by taking a weighted average of the features of each cluster;

[0039] S6007: Repeat steps S6005 and S6006 until the log-likelihood of the model converges, that is, the parameter change is less than the set threshold ∈, and the establishment of the Gaussian mixture universal background model is completed.

[0040] After adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art. Of course, any product implementing the present invention does not necessarily need to achieve all the advantages described below at the same time:

[0041] The present invention can automatically extract key local spectral features and temporal dependencies from dialect speech signals, avoiding the manual reliance on feature extraction in traditional methods, thereby significantly improving the recognition accuracy of dialects. By combining convolutional neural networks (CNNs) and bidirectional gated recurrent units (BiGRUs), the present invention can deeply explore local details and long-term dependency features in speech signals, ensuring that the model accurately captures complex speech features. In addition, the present invention effectively combines features at different levels through multi-level feature fusion, further improving the model's ability to recognize pronunciation differences between dialects. When faced with diverse dialect data, the invention exhibits excellent robustness and adaptability, and can effectively deal with complex factors such as pronunciation differences and voice changes, thereby achieving higher accuracy and stability in dialect classification tasks.

[0042] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The drawings described below are only some embodiments. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0044] In the figure:

[0045] Figure 1 It is a structural diagram.

[0046] It should be noted that these drawings and textual descriptions are not intended to limit the conceptual scope of the present invention in any way, but are intended to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0047] The present invention will now be described in further detail with reference to the accompanying drawings.

[0048] See also Figure 1 As shown, in this embodiment, a dialect recognition method based on a language model is provided, comprising the following steps

[0049] S1: Preprocess the input dialect speech signal to obtain time-frequency features;

[0050] S2: inputting the time-frequency features into an empirical mode decomposition (EMD) module, decomposing the dialect speech signal into multiple empirical mode components (IMFs) and extracting the logarithmic energy features of each IMF;

[0051] S3: Fusing the logarithmic energy feature extracted in step S2 with the time-frequency feature extracted in step S1 to obtain a multi-dimensional feature vector for subsequent deep learning model training;

[0052] S4: input the fused features obtained in step S3 into a convolutional neural network (CNN), and extract the local spectrum features of the dialect signal by using a convolution operation;

[0053] S5: Input the local spectrum features extracted by CNN in step S4 into a bidirectional gated recurrent unit (BiGRU) module to extract the temporal features of the dialect speech;

[0054] S6: Input the temporal features extracted by BiGRU in step S5 into the Gaussian mixture general background model to classify and identify the dialect;

[0055] S7: Determine the dialect category based on the log-likelihood probability difference between the improved GMM-UBM model and the background model (UBM) in step S6, and output the final recognition result.

[0056] In this embodiment, the time-frequency feature is extracted by the Mel-frequency cepstral coefficient method to obtain the feature vector F of each frame of speech. t, where t represents the time frame index, And d is the dimension of the feature vector of each frame.

[0057] In this embodiment, the steps of inputting the time-frequency features into an empirical mode decomposition (EMD) module, decomposing the dialect speech signal into multiple empirical mode components (IMFs) and extracting the logarithmic energy features of each IMF are as follows:

[0058] S201: Determine the local maximum and local minimum of the original speech signal X(t), fit the upper and lower envelopes by cubic spline interpolation, obtain the upper envelope U(t) and the lower envelope L(t), and then, by formula Calculate the mean curve m1(t) of the envelope line, and then calculate the empirical mode component h1(t) according to the obtained mean curve m1(t) by the formula h1(t)=X(t)-m1(t);

[0059] S202: If h1(t) does not meet the preset threshold standard (for example, the standard deviation is less than 0.2), repeat step S201 to obtain a new empirical mode component until the component meets the standard;

[0060] S203: Extracting logarithmic energy features E from all empirical mode components IMF (t), where the logarithmic energy feature is calculated as: Among them, h i (t) represents the i-th empirical mode component.

[0061] In this embodiment, the fusion feature obtained in step S3 is input into a convolutional neural network (CNN), and the steps of extracting the local spectrum feature of the dialect signal by using the convolution operation are as follows:

[0062] S401: Input the feature vector F obtained in step S3 and pass it to the convolution layer. The convolution operation extracts local features through the kernel function K to obtain the feature map F CNN , whose expression is: Among them, F i represents the i-th element in the input feature vector, K i is the convolution kernel, n is the size of the convolution kernel;

[0063] S402: Downsample the convolutional feature map through a pooling layer, usually using a maximum pooling operation, to obtain a feature map P CNN Then, the pooled feature map is flattened to obtain the vector P CNN-flat , so as to be input into the subsequent fully connected layer.

[0064] In this embodiment, the local spectrum features extracted by CNN in step S4 are input into a bidirectional gated recurrent unit (BiGRU) module, and the steps of extracting the temporal features of the dialect speech are as follows:

[0065] S501: Input feature vector P CNN-flat The signal is passed to the bidirectional gated recurrent unit (BiGRU), which processes the signal forward and backward at the same time, obtains the front and back time sequence information of the dialect speech, and outputs the time sequence feature F BiGRU , the calculation process of BiGRU is: GRU forward =σ(W f F input +b f ), GRU backward =σ(W b F input +b b ) where σ represents the activation function, W f and W b are the weight matrices of the forward and backward GRUs, respectively, input is the input feature, b f and b b is the bias term.

[0066] S502: Combine the forward and backward GRU outputs obtained in step S501 to obtain the bidirectional temporal feature F BiGRU , which is used for subsequent classification decisions.

[0067] In this embodiment, the steps of inputting the temporal features extracted by BiGRU in step S5 into the Gaussian mixture general background model to classify and identify the dialect are as follows:

[0068] S601: Using a large-scale dialect dataset to train a universal background model (UBM) to obtain a Gaussian mixture model (GMM);

[0069] S602: Adaptively adjust the UBM using the Maximum A Posteriori Adaptive (MAP) algorithm to obtain a model λ suitable for the target dialect target ; The system has multiple dialect models. We first train a universal background model (UBM) using data from multiple dialects. Then, when the system needs to recognize Cantonese, we will use the speech data of Cantonese to adjust the UBM and obtain a target dialect model for Cantonese; similarly, when the system recognizes other dialects, it will also use speech data from different dialects for training, so as to obtain target dialect models that adapt to the characteristics of these dialects.

[0070] S603: By screening the training samples, the dialect model is optimized and its discrimination is enhanced, and finally a dialect model with stronger adaptability is obtained.

[0071] In this embodiment, dialect classification is performed based on the log-likelihood probability difference between the improved GMM-UBM model and the background model (UBM) in step S6, and its expression is: i (X) = logp(X|λ target )-logp(X|λ UBM ) where λ target is the target dialect model, λ UBM is the background model, and X is the input feature vector.

[0072] Calculate the log-likelihood Λ for each class i i (X), and determine the dialect category by maximizing the log-likelihood value. Finally, the log-likelihood probability difference is used to select the dialect category with the largest score and output the final recognition result.

[0073] In this embodiment, the steps of establishing the Gaussian mixture general background model are:

[0074] S6001: The time-frequency feature data {F1, F2, ..., F T} is input into the Gaussian mixture model for modeling. First, the mean vector μ and covariance matrix ∑ of the feature data are calculated.

[0075] S6002: Use the expectation maximization (EM) algorithm to train the GMM and estimate the parameters of the GMM by maximizing the likelihood function. The probability density function of the GMM is: Where λ = {π i , μ i ,∑ i} represents the parameters of GMM, π i is the weight of the ith Gaussian distribution, μ i and∑ i are the mean vector and covariance matrix of the i-th Gaussian distribution respectively.

[0076] S6003: Initialize GMM parameter π i , μ i ,∑ i Use the K-means clustering algorithm to perform preliminary clustering on the input features, and obtain the mean and covariance matrix of each cluster as the initial parameters of GMM;

[0077] S6004: Perform K-means clustering on the data set and calculate the mean of each cluster and the covariance matrix As initial GMM parameters;

[0078] S6005: Execute the EM algorithm to estimate parameters and calculate the responsibility γ through the expected step it , whose expression is: Among them, γit Indicates the responsibility of the i-th Gaussian component to the t-th feature;

[0079] S6006: Update the parameters of GMM according to the degree of responsibility in step S6005, and update the weight π i , mean μ i , and the covariance matrix ∑ i , update the parameters by taking a weighted average of the features of each cluster;

[0080] S6007: Repeat steps S6005 and S6006 until the log-likelihood of the model converges, that is, the parameter change is less than the set threshold ∈, and the establishment of the Gaussian mixture universal background model is completed.

[0081] The trained GMM model can be used as a universal background model (UBM) for subsequent target dialect model training. The target dialect model is trained using the MAP adaptive algorithm to adjust the large-scale universal background model parameters to the characteristics of a specific dialect. The convolutional neural network (CNN) of the present invention is used to extract local features, BiGRU is used for time series modeling, and finally classification is performed through GMM-UBM. During training, CNN uses 4 convolutional layers and the number of BiGRU hidden units is 128.

[0082] Comparative Example 1: Traditional GMM-UBM dialect recognition method,The GMM-UBM method is a traditional dialect recognition technology, which models the common features of all dialects by training a universal background model (UBM). Then, the maximum a posteriori adaptive (MAP) algorithm is used to adjust the background model to obtain a target model suitable for a specific dialect. This method mainly relies on the Gaussian mixture model to estimate the feature distribution of the dialect.

[0083] Comparative Example 2: Dialect recognition method based on convolutional neural network (CNN) + long short-term memory network (LSTM). This method uses a deep learning model. First, the convolutional neural network (CNN) is used to extract the local features of the dialect speech. Then the long short-term memory network (LSTM) is used to capture the temporal features of the speech. Finally, the full connection layer is used for classification. This method does not require an explicit feature extraction step and automatically learns the characteristics of the dialect from the data.

[0084] Test results of iFLYTEK dialect dataset

[0085] The test was conducted using the dataset from the 2018 iFLYTEK Dialect Identification AI Challenge. The dataset includes 10 dialects: Changsha, Hebei, Hefei, Hakka, Minnan, Nanchang, Ningxia, Shaanxi, Shanghai, and Sichuan. The dataset for each dialect was recorded by 30 people. The training set for each dialect has 6,000 voices, and the test set has 500 voices. Experimental analysis was conducted on this dataset.

[0086] Experimental environment: All models were trained and tested using the same NVIDIA Tesla V100 GPU computer. All models were implemented using the TensorFlow framework to ensure consistency of the experiment.

[0087] Training parameters

[0088] GMM-UBM method: The EM algorithm is used to train GMM, MAP is adaptively adjusted, 16 Gaussian components are used for modeling, and the number of iterations is 100.

[0089] CNN+LSTM method: The number of CNN convolution layers is set to 4, the size of each convolution kernel is 3x3, the number of LSTM layers is set to 2, the number of hidden units is 128, the optimizer is Adam, and the learning rate is 0.001.

[0090] Dialect recognition method based on language model (the present invention): Convolutional neural network (CNN) is used to extract local features, BiGRU is used for time series modeling, and finally classification is performed through GMM-UBM. During training, CNN uses 4 convolutional layers and the number of BiGRU hidden units is 128.

[0091] Table 1: Comparison of accuracy rates of different dialects

[0092]

[0093] It can be seen from the table that the accuracy of the present invention (CNN+BiGRU method) in each dialect is higher than that of the traditional GMM-UBM method and CNN+LSTM method, especially in some dialects (such as Shanghai and Hakka), the improvement of accuracy is particularly significant.

[0094] It can be seen from the experimental results that the accuracy of the present invention (CNN+BiGRU method) in each dialect is higher than that of the traditional GMM-UBM method and CNN+LSTM method, especially in some dialects (such as Shanghai and Hakka), the improvement in accuracy is particularly significant. This is mainly due to the advantages of the present invention: first, the convolutional neural network (CNN) can effectively extract local spectral features in dialect speech signals, enhancing the expressiveness of features; second, the bidirectional gated recurrent unit (BiGRU) can simultaneously capture the forward and backward temporal dependencies of speech signals, improving the modeling ability of temporal information; third, combined with the GMM-UBM model, dialect classification is more accurate, especially in complex dialects, showing stronger robustness and adaptability. Through end-to-end training of deep learning, the present invention avoids the limitations of manual feature design in traditional methods, and can more comprehensively learn the unique characteristics of dialects, thereby achieving higher accuracy in dialect recognition tasks.

[0095] The present invention is not limited to the above-mentioned embodiments. Anyone should be aware that any structural changes made under the enlightenment of the present invention, and any technical solutions that are the same or similar to the present invention, fall within the protection scope of the present invention. The technology, shape, and structural parts not described in detail in the present invention are all well-known technologies.

Claims

1. A dialect recognition method based on a language model, characterized in that: The following steps are involved: S1: Preprocess the input dialect speech signal to obtain time-frequency features; S2: inputting the time-frequency features into an empirical mode decomposition module, decomposing the dialect speech signal into multiple empirical mode components and extracting the logarithmic energy features of each IMF; S3: Fusing the logarithmic energy feature extracted in step S2 with the time-frequency feature extracted in step S1 to obtain a multi-dimensional feature vector for subsequent deep learning model training; S4: input the fusion features obtained in step S3 into the convolutional neural network, and extract the local spectrum features of the dialect signal by using the convolution operation; S5: Input the local spectrum features extracted by CNN in step S4 into a bidirectional gated recurrent unit (BiGRU) module to extract the temporal features of the dialect speech; S6: Input the temporal features extracted by BiGRU in step S5 into the Gaussian mixture general background model to classify and identify the dialect; S7: Determine the dialect category based on the log-likelihood probability difference between the improved Gaussian mixture general background model and the background model in step S6, and output the final recognition result.

2. The dialect identification method based on a language model according to claim 1, characterized in that: The time-frequency features are extracted by the Mel-frequency cepstral coefficient method to obtain the feature vector F of each frame of speech. t , where t represents the time frame index, And d is the dimension of the feature vector of each frame.

3. The dialect identification method based on a language model according to claim 1, characterized in that: The time-frequency features are input into the empirical mode decomposition module, the dialect speech signal is decomposed into multiple empirical mode components and the logarithmic energy features of each IMF are extracted as follows: S201: Determine the local maximum and local minimum of the original speech signal X(t), fit the upper and lower envelopes by cubic spline interpolation, obtain the upper envelope U(t) and the lower envelope L(t), and then, by formula Calculate the mean curve m1(t) of the envelope line, and then calculate the empirical mode component h1(t) according to the obtained mean curve m1(t) by the formula h1(t)=X(t)-m1(t); S202: If h1(t) does not meet the preset threshold standard, repeat step S201 to obtain a new empirical mode component until the component meets the standard; S203: Extracting logarithmic energy features E from all empirical mode components IMF (t), where the logarithmic energy feature is calculated as: Among them, h i (t) represents the i-th empirical mode component.

4. The dialect identification method based on a language model according to claim 1, characterized in that: The fusion features obtained in step S3 are input into the convolutional neural network, and the steps of extracting the local spectrum features of the dialect signal by using the convolution operation are as follows: S401: Input the feature vector F obtained in step S3 and pass it to the convolution layer. The convolution operation extracts local features through the kernel function K to obtain the feature map F CNN , whose expression is: Among them, F i represents the i-th element in the input feature vector, K i is the convolution kernel, and Tuo is the size of the convolution kernel; S402: Downsample the convolutional feature map through a pooling layer, usually using a maximum pooling operation, to obtain a feature map P CNN Then, the pooled feature map is flattened to obtain the vector P CNN-flat , so as to be input into the subsequent fully connected layer.

5. The dialect identification method based on language model according to claim 1, characterized in that: The local spectrum features extracted by CNN in step S4 are input into the Bidirectional Gated Recurrent Unit (BiGRU) module to extract the temporal features of the dialect speech as follows: S501: Input feature vector P CNN-flat The signal is passed to the bidirectional gated recurrent unit, and the BiGRU processes the signal forward and backward at the same time, obtains the before and after time sequence information of the dialect speech, and outputs the time sequence feature F BiGRU , the calculation process of BiGRU is: GRU forward =σ(W f F input +b f ), GRU backward =σ(W b F input +b b ) where σ represents the activation function, W f and W b are the weight matrices of the forward and backward GRUs, respectively, input is the input feature, b f and b b is the bias term; S502: Combine the forward and backward GRU outputs obtained in step S501 to obtain the bidirectional temporal feature F BiGRU , which is used for subsequent classification decisions.

6. The dialect identification method based on language model according to claim 1, characterized in that: The steps of inputting the temporal features extracted by BiGRU in step S5 into the Gaussian mixture general background model to classify and identify the dialect are as follows: S601: Use a large-scale dialect data set to train a general background model to obtain a Gaussian mixture model; S602: Adaptively adjust the UBM using the maximum a posteriori adaptive algorithm to obtain a model λ suitable for the target dialect target ; S603: By screening the training samples, the dialect model is optimized and its discrimination is enhanced, and finally a dialect model with stronger adaptability is obtained.

7. The dialect identification method based on language model according to claim 1, characterized in that: Dialect classification is performed based on the log-likelihood probability difference between the Gaussian mixture general background model and the background model in step S6, and its expression is: i (X) = logp(X|λ target )-logp(X|λ UBM )in, λ target is the target dialect model, λ UBM is the background model, X is the input feature vector, and finally, the log-likelihood value A of each category i is calculated. i (X), and determine the dialect category by maximizing the log-likelihood value. Finally, the log-likelihood probability difference is used to select the dialect category with the largest score and output the final recognition result.

8. The dialect identification method based on a language model according to claim 1, characterized in that: The steps to build a Gaussian mixture general background model are: S6001: The time-frequency feature data {F1, F2, ..., F T } Input into Gaussian mixture model for modeling, first calculate the mean vector μ and covariance matrix ∑ of the feature data; S6002: GMM is trained using the expectation maximization algorithm, and the parameters of GMM are estimated by maximizing the likelihood function. The probability density function of GMM is: Where λ = {π i , μ i ,∑ i } represents the parameters of GMM, π i is the weight of the ith Gaussian distribution, μ i and∑ i are the mean vector and covariance matrix of the i-th Gaussian distribution respectively; S6003: Initialize GMM parameter π i , μ i ,∑ i Use the K-means clustering algorithm to perform preliminary clustering on the input features, and obtain the mean and covariance matrix of each cluster as the initial parameters of GMM; S6004: Perform K-means clustering on the data set and calculate the mean of each cluster and the covariance matrix As initial GMM parameters; S6005: Execute the EM algorithm to estimate parameters and calculate the responsibility γ through the expected step it , whose expression is: Among them, γ it Indicates the responsibility of the i-th Gaussian component to the t-th feature; S6006: Update the parameters of GMM according to the degree of responsibility in step S6005, and update the weight π i , mean μ i , and the covariance matrix ∑ i , update the parameters by taking a weighted average of the features of each cluster; S6007: Repeat steps S6005 and S6006 until the log-likelihood of the model converges, that is, the parameter change is less than the set threshold ∈, and the establishment of the Gaussian mixture universal background model is completed.

Images