Agricultural disease and pest dialect voice intelligent recognition method based on deep learning technology

Abstract

The invention discloses an agricultural disease and pest dialect speech intelligent recognition method based on a deep learning technology, and the method comprises the steps: combining a Mel-frequency cepstral coefficient with a convolutional neural network, and dividing a whole process into two stages, the pre-training stage comprises the steps of recording and storing a long voice file, extracting the characteristics of voice signals in the file and training a neural network model, and the retraining stage comprises the steps of recording and storing a short voice file, extracting the characteristics of the voice signals in the file, verifying a test result, adjusting model parameters and storing an optimal model; the dialect voice information reported by monitoring personnel in an agricultural field is intelligently recognized by considering accent of people in different regions, so that the recognition accuracy of complex dialect voice signals is effectively improved, and the recognition method is particularly suitable for understanding related terms of agricultural diseases and insect pests.

Description

Technical Field

[0001] The present invention belongs to the field of information processing technology, and specifically is a method for intelligent recognition of agricultural pest and disease dialect speech based on deep learning technology. Background Art

[0002] The diversity of dialects has brought new challenges to speech recognition. Especially in the agricultural field, if pests and diseases can be described and recorded using voice input based on intelligent recognition technology, this method can greatly improve the accuracy and efficiency of pest and disease management, help agricultural practitioners respond in time, and reduce the impact of pests and diseases on crops. However, agricultural practitioners often use local dialects and specific terms when describing pests and diseases. Dialects in different regions have significant differences in pronunciation, vocabulary and grammar, which affects the adaptability and accuracy of speech recognition models in the agricultural field. Summary of the invention

[0003] Purpose of the invention: In order to solve the problem that existing speech recognition methods are not applicable to local dialects and specific terms used to describe pests and diseases, the present invention proposes an intelligent recognition method for agricultural pest and disease dialect speech based on deep learning technology. The present invention focuses on the recognition of dialect speech and uses deep learning models for training and optimization, which greatly enhances the understanding and recognition of specific dialect features, can greatly improve the efficiency of agricultural practitioners in dealing with pests and diseases, promote the sustainable development of agriculture, and provide precise technical support for modern agricultural management.

[0004] Technical solution: A method for intelligent recognition of dialect speech of agricultural pests and diseases based on deep learning technology, comprising the following steps: Step 1: Obtain dialect speech data, annotate, segment and denoise the dialect speech data to obtain dialect speech samples; process each dialect speech sample to obtain the corresponding MFCC features, thereby forming an MFCC feature data set, and each MFCC feature represents a dialect speech sample; according to the number of text labels contained in a dialect speech sample, the MFCC feature data set is divided into an MFCC feature data set corresponding to a long dialect speech sample and an MFCC feature data set corresponding to a short dialect speech sample; Step 2: The MFCC feature data set corresponding to the long dialect speech sample is used as the long dialect speech sample training set, and the convolutional neural network model is trained with the long dialect speech sample training set. The parameters of the convolutional neural network model are updated through the optimizer to obtain a pre-trained convolutional neural network model that meets the performance requirements; Step 3: Divide the MFCC feature data set corresponding to the short dialect speech sample into a short dialect speech sample training set and a short dialect speech sample test set. Use the short dialect speech sample training set to retrain the pre-trained convolutional neural network model. Use the short dialect speech sample test set to test the performance of the retrained convolutional neural network model, and finally obtain the dialect speech recognition model. Step 4: Process the dialect speech data input by the user to obtain the corresponding MFCC features, input the MFCC features into the dialect speech recognition model, obtain the dialect speech recognition results, and use the dialect speech recognition results as the text labels of its MFCC features and add them to the MFCC feature data set corresponding to the short dialect speech sample; Step 5: At set intervals, the pre-trained convolutional neural network model is retrained using the MFCC feature data set corresponding to the latest short dialect speech sample to obtain the latest dialect speech recognition model.

[0005] Furthermore, the obtaining of the corresponding MFCC features specifically includes: Preprocessing the dialect speech sample / dialect speech data input by the user to obtain a pre-emphasized signal; Perform frame segmentation on the pre-emphasized signal to obtain multiple frame signals, where adjacent frame signals overlap; Perform windowing processing on each frame of signal; Perform fast Fourier transform on each frame of the windowed signal to convert the time domain signal into a frequency domain signal; Design a Mel filter bank, apply the Mel filter bank to each frequency domain signal, and extract the energy of the Mel frequency band; According to the energy of the Mel frequency band, the discrete cosine transform is calculated to obtain the corresponding MFCC features; The corresponding MFCC features are normalized to obtain the corresponding MFCC features.

[0006] Further, the dialect speech sample / the dialect speech data input by the user is preprocessed to obtain a pre-emphasized signal, specifically including: The dialect speech sample / the dialect speech data input by the user is pre-emphasized by a first-order high-pass filter, and the first-order high-pass filter is expressed as: , where represents the pre-emphasis coefficient, Represents the signal after pre-emphasis, Represents dialect speech samples / dialect speech data input by the user, express The value of the previous sampling point, Indicates the sampling point number.

[0007] Furthermore, the windowing process is performed on each frame of signal, and the specific operations include: Use the following window function to perform windowing on each frame of signal: ; in, is the length of the window.

[0008] Furthermore, the fast Fourier transform is performed on each frame of the windowed signal to convert the time domain signal into a frequency domain signal, which is expressed as: ; In the formula, represents the window function, represents the index of the frequency component in the frequency domain, is the Fourier basis function, Represents an imaginary unit.

[0009] Furthermore, the Mel scale in the Mel filter bank is expressed as: ; In the formula, Indicates the signal frequency.

[0010] Furthermore, for each frequency domain signal, a Mel filter bank is applied to extract the energy of the Mel frequency band, which is expressed as: ; in, represents the response of the mth Mel filter.

[0011] Furthermore, the discrete cosine transform is calculated based on the energy of the Mel frequency band to obtain the corresponding MFCC features, which is expressed as: ; in, represents MFCC features, and M represents the number of Mel filters in the Mel filter bank.

[0012] Furthermore, the cross entropy loss function is used in the process of training the convolutional neural network model using the long dialect speech sample training set, which is expressed as: ; in, is the true text label of the d-th long dialect speech sample, is the predicted probability of the d-th long dialect speech sample; D is the number of samples in the long dialect speech sample training set; The parameters of the convolutional neural network model are updated by the optimizer, which is expressed as: ; in, is the first-order moment estimate of the gradient, is the second-order moment estimate of the gradient, represents a constant, represents the learning rate, Represents the parameters of the convolutional neural network model in the previous iteration.

[0013] Furthermore, in the process of retraining the pre-trained convolutional neural network model using the short dialect speech sample training set, the optimizer is used to update the model parameters, which is expressed as: ,in, Represents the parameters of the convolutional neural network model of the current round iteration, Represents the loss function About parameters gradient.

[0014] Beneficial effects: Compared with the prior art, the present invention has the following advantages: (1) The method of the present invention combines Mel-frequency cepstral coefficients (MFCC) and convolutional neural networks (CNN), and divides the entire process into two stages. The pre-training stage includes recording and storing long voice files, extracting features of voice signals in long voice files, and training neural network models. The retraining stage includes recording and storing short voice files, extracting features of voice signals in short voice files, verifying test results, adjusting model parameters, and saving the optimal model. By considering the accents of people in different regions, all voices are processed into MFCC features. Based on the model, the existing same dialect voice is trained at the beginning. When the same dialect voice is encountered again, the obtained MFCC features are passed into the deep learning neural network of the model for prediction, which can accurately identify the dialect information. In this way, the dialect voice information reported orally by agricultural practitioners can be intelligently recognized, which effectively improves the recognition accuracy of complex dialect voice signals, especially in the understanding of agricultural pests and diseases related terms. (2) The method of the present invention describes and records pests and diseases through voice input, which greatly improves the efficiency of agricultural practitioners in recording and managing pests and diseases, and can make corresponding solutions more timely; (3) The method of the present invention provides an intelligent solution for agricultural pest management, promotes the digital transformation of agriculture, and improves production efficiency; (4) The method of the present invention greatly improves dialect speech recognition, thereby increasing the coverage of information acquisition, reducing the possibility of missing key information due to language barriers, greatly improving pest and disease identification, and more accurate information can further improve the accuracy of related management measures, promoting the development of sustainable agriculture and helping to protect the ecological environment; (5) The method of the present invention is not only applicable to agricultural pest and disease identification, but can also be extended to dialect speech recognition and intelligent monitoring in other fields, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flowchart of a method for intelligent recognition of dialect speech of agricultural pests and diseases based on deep learning technology proposed by the present invention; Figure 2 Schematic diagram of the convolutional neural network (CNN) architecture of the present invention. DETAILED DESCRIPTION

[0016] The technical solution of the present invention is now further described in conjunction with the accompanying drawings and embodiments.

[0017] Example: Many agricultural practitioners use dialects for daily communication, and they may not be able to use Mandarin to accurately express the pest and disease information they observe in the field. If the dialect information they speak can be collected in time through the spectrum graph, and then the deep learning neural network is used for intelligent speech recognition, the coverage of information acquisition can be greatly improved, and the possibility of missing key information due to language barriers can be reduced.

[0018] Therefore, if Figure 1 As shown, this embodiment proposes an intelligent recognition method for dialect speech of agricultural pests and diseases based on deep learning technology, focusing on the recognition of dialect speech, using deep learning models for training and optimization, enhancing the understanding and recognition effect of specific dialect features, thereby promoting the development of smart agriculture, improving the efficiency of agricultural practitioners in dealing with pests and diseases, promoting the sustainable development of agriculture, and providing precise technical support for modern agricultural management. It mainly includes the following steps: Step 1: In the context of agricultural pests and diseases, specific vocabulary and expressions need to be integrated through professional knowledge. Therefore, it is particularly important to develop high-quality speech data sets and annotations. This requires that these data sets should include diverse dialect samples that can fully reflect the descriptions and countermeasures of agricultural practitioners in different regions on pests and diseases; this step constructs a dialect speech data set specifically for agricultural pests and diseases, covering dialect samples from different regions to ensure the diversity and representativeness of the data; specifically, dialect speech data for agricultural pests and diseases are collected, and each dialect speech data is annotated, segmented and denoised to obtain dialect speech samples; in this embodiment, the annotation is combined with the knowledge of agricultural experts, which can not only improve the performance of the model, but also enhance its reliability in practical applications. Each dialect speech sample is processed to obtain the corresponding MFCC feature, thereby forming an MFCC feature data set, in which an MFCC feature represents a dialect speech sample, which is used for the preliminary learning of the pre-training model on the dialect spectrogram. In this embodiment, according to the number of text labels contained in each dialect speech sample, the dialect speech samples are divided into long dialect speech samples and short dialect speech samples. Therefore, correspondingly, the MFCC feature data set is divided into an MFCC feature data set corresponding to the long dialect speech sample and an MFCC feature data set corresponding to the short dialect speech sample. In this embodiment, the duration of each text label being read out in dialect is 2 seconds.

[0019] For each dialect speech sample, the spectrum graph is different, so for each dialect speech sample, the MFCC features of the dialect speech sample can be obtained by preprocessing, frame segmentation, application of window function, fast Fourier transform, Mel filter, logarithmic energy calculation, discrete cosine transform and feature normalization, and the audio data can be converted into a feature matrix; the specific operations include: Load audio data; Pre-processing: Audio data usually needs to be pre-emphasized to enhance high-frequency components. Pre-emphasis can be accomplished by a simple first-order high-pass filter, which is expressed as: , where Indicates the pre-emphasis coefficient, usually between 0.95 and 0.98. Represents the pre-emphasized signal, that is, the output signal after filtering, represents the audio data input to the first-order high-pass filter, which represents the dialect speech sample in this step and the dialect speech data input by the user in normal use. express The value of the previous sampling point, Represents the index of discrete time points, that is, the sampling point number of audio data. Normalize the output signal after filtering.

[0020] Frame segmentation: Audio data is a time-varying signal. In order to process the signal, it is usually divided into short time frames. The length of each frame is usually between 20 and 40 milliseconds, and there will be a certain overlap between frames (usually 50%). Apply window function: In order to reduce the edge effect caused by framing, the signal of each frame is windowed. The commonly used window function is the Hamming window, and its formula is: ; in, is the length of the window.

[0021] Fast Fourier Transform: Perform a fast Fourier transform on each frame of the signal to convert the time domain signal into a frequency domain signal. This step generates spectrum information, which represents the intensity of each frequency component in the signal. The formula is: ; In the formula, represents the window function, Represents the index of the frequency component in the frequency domain, corresponding to a point on the discrete frequency axis, used to represent a frequency component of the signal. is the Fourier basis function, which is used to transform the time domain signal into the frequency domain and decompose the signal into sine waves and cosine waves of different frequencies. Represents an imaginary unit.

[0022] Mel filtering: Apply the Mel filter bank to convert the spectrum to the Mel frequency scale. The design of the Mel scale helps define the distribution and frequency center of the Mel filter bank on the Mel frequency scale, thereby The spectrum signal is calculated in Weighted, extract the energy of the Mel frequency band, the Mel scale is expressed as: ; In the formula, Indicates the signal frequency.

[0023] Logarithmic energy calculation: Calculate the logarithmic energy of each Mel filter output, that is, the energy of the Mel frequency band, expressed as: ; in, represents the response of the mth Mel filter.

[0024] Discrete cosine transform (DCT): Based on the energy of the Mel frequency band, the discrete cosine transform is calculated to obtain the corresponding MFCC features, which are expressed as: ; in, represents MFCC features, and M represents the number of Mel filters in the Mel filter bank.

[0025] Feature normalization: Normalize the MFCC features.

[0026] Step 2: Slice the extracted MFCC features to meet the requirements of model input; Step 3: Build a convolutional neural network (CNN) model, use the MFCC feature data set corresponding to the long dialect speech sample as the long dialect speech sample training set, train the convolutional neural network model with the long dialect speech sample training set, update the parameters of the convolutional neural network model through the optimizer, and obtain a pre-trained convolutional neural network model that meets the performance requirements.

[0027] This embodiment combines Mel-frequency cepstral coefficients (MFCC) with convolutional neural networks to improve the recognition ability of complex dialect speech signals, especially for terms and descriptions related to pests and diseases.

[0028] like Figure 2 As shown, the convolutional neural network models used in this embodiment are: Input layer: The input shape is (20, 11, 1), which means that the input data is a two-dimensional feature map with 20 rows, 11 columns and 1 channel. This input layer accepts the MFCC feature matrix of the preprocessed speech.

[0029] Convolution layer: The convolution operation uses 64 filters, each of which is 2×2 in size. The activation function uses ReLU activation to introduce nonlinearity to the model, thereby improving the ability to express complex features. The output shape will be reduced by the convolution operation, and the specific shape needs to be calculated based on the filling method.

[0030] Pooling layer (MaxPooling2D): Use a 2×2 window to perform the maximum pooling operation. This operation effectively reduces the size of the feature map by taking the maximum value of the local area of ​​the feature map, thereby reducing the computational complexity and preventing overfitting. The output shape will continue to be reduced to 1 / 4 of the original (the width and height are half of the original).

[0031] The first Dropout layer: Set the dropout rate to 0.25, that is, randomly drop the output of some neurons with a probability of 25%. The function of Dropout is to prevent overfitting and improve the generalization ability of the model.

[0032] Flattening layer: Flattens the multi-dimensional feature map into a one-dimensional vector to facilitate input to the subsequent fully connected layer. This step integrates the spatial features extracted by the convolution and pooling layers.

[0033] The first fully connected layer: contains 128 neurons and uses the ReLU activation function. This layer further captures the high-level patterns of the input features through linear transformation and nonlinear activation.

[0034] Second Dropout layer: Set the dropout rate rate=0.25 again to further prevent overfitting.

[0035] The second fully connected layer: contains the same number of neurons as the number of categories (num_classes), and outputs the probability distribution of each category. The activation function is softmax, which is used for multi-classification problems to probabilize the output and ensure that the sum of the probabilities of all categories is 1. This CNN model extracts the key features of the input through convolution and pooling layers, and then integrates and classifies the features through the fully connected layer. At the same time, the Dropout technology is used to effectively suppress overfitting. The overall structure is compact and suitable for processing classification tasks after speech feature extraction.

[0036] The following calculations are performed in the constructed convolutional neural network model: In the convolutional layer, for the input MFCC features and convolution kernel The output of the convolutional layer is calculated as follows: ; In the formula, , Represents the matrices of the convolutional layer outputs The row and column indices in represent the location of the feature map produced after the convolution operation. , Represent the convolution kernels The row and column indices of the elements in represent the position of each element of the convolution kernel. Represents the size of the convolution kernel.

[0037] For the maximum pooling operation, the formula is: ; In the formula, represents the pixel value within the pooling window, , represents the offset of the pooling window, Represents the maximum value operation, that is, the maximum value in the current pooling window is selected as the result.

[0038] The cross entropy loss function is used in the process of training the convolutional neural network model using the long dialect speech sample training set, which is expressed as: ; in, is the true text label of the d-th long dialect speech sample, is the predicted probability of the d-th long dialect speech sample; D is the number of samples in the long dialect speech sample training set.

[0039] The Adam optimizer is used to update the parameters during the training process, which is expressed as: ; in, is the first-order moment estimate of the gradient, is the second-order moment estimate of the gradient, represents a constant, in this embodiment, a very small constant used to prevent division by zero errors, Represents the learning rate, which is a hyperparameter that controls the update amplitude during the optimization process. Represents the parameters of the convolutional neural network model in the previous iteration.

[0040] In the model prediction phase, the predicted probability of each category is calculated ,in Indicates the label, Represents input; sorts the predicted probabilities from high to low and outputs them.

[0041] Step 4: Divide the MFCC feature set corresponding to the short dialect speech sample into a short dialect speech sample training set and a short dialect speech sample test set, and use the short dialect speech sample training set to retrain the pre-trained convolutional neural network model; use the short dialect speech sample test set to test the performance of the retrained convolutional neural network model, and finally obtain the dialect speech recognition model; Step 5: Entering the test phase, the user records the symptoms of pests and diseases through voice input, and MFCC features are extracted from the voice data. The extracted MFCC features are input into the dialect speech recognition model to obtain the dialect speech recognition results; Step 6: Ask the user whether the dialect speech recognition result is accurate. If it is accurate, continue to predict the next test speech. Otherwise, further optimize the parameters of the dialect speech recognition model. During the test, regardless of whether the dialect speech recognition result is accurate or not, the short dialect speech sample training set is updated according to the real label given by the user; Step 7: Use the updated short dialect speech sample training set to retrain the dialect speech recognition model, and the loss function still uses the cross entropy loss.

[0042] Use the optimizer to update the model parameters, expressed as: ; in, Represents the parameters of the convolutional neural network model of the current round iteration, Represents the loss function About parameters gradient.

[0043] like Figure 1 As shown, when users use the system, they record the symptoms of pests and diseases through voice input. This natural way of communication enables them to express complex agricultural problems more easily, especially when using dialects, and can more accurately describe the local pest and disease phenomena. Users are not only providers of information, but also executors of decisions. After receiving the pest and disease identification results provided by the system, they need to combine their own experience and actual conditions, evaluate the accuracy of the results, and take timely measures. In addition, users may encounter various technical difficulties during use, so their feedback is crucial to the improvement and optimization of the system, mainly because the convolutional neural network model is retrained and optimized by continuously receiving feedback and new data from agricultural practitioners, thereby improving recognition accuracy and ensuring its effectiveness and reliability in practical applications. The users of this embodiment are usually agricultural producers with rich practical experience, who understand the pest and disease characteristics and prevention and control measures of the crops they plant.

Claims

1. A method for intelligent recognition of agricultural pests and diseases dialect speech based on deep learning technology, characterized by: The following steps are involved: Step 1: Obtain dialect speech data, annotate, segment and denoise the dialect speech data, and obtain dialect speech samples; Each dialect speech sample is processed to obtain the corresponding MFCC features, thereby forming an MFCC feature data set, where each MFCC feature represents a dialect speech sample; according to the number of text labels contained in a dialect speech sample, the MFCC feature data set is divided into an MFCC feature data set corresponding to a long dialect speech sample and an MFCC feature data set corresponding to a short dialect speech sample; Step 2: The MFCC feature data set corresponding to the long dialect speech sample is used as the long dialect speech sample training set, and the convolutional neural network model is trained with the long dialect speech sample training set. The parameters of the convolutional neural network model are updated through the optimizer to obtain a pre-trained convolutional neural network model that meets the performance requirements; Step 3: Divide the MFCC feature data set corresponding to the short dialect speech sample into a short dialect speech sample training set and a short dialect speech sample test set. Use the short dialect speech sample training set to retrain the pre-trained convolutional neural network model. Use the short dialect speech sample test set to test the performance of the retrained convolutional neural network model, and finally obtain the dialect speech recognition model. Step 4: Process the dialect speech data input by the user to obtain the corresponding MFCC features, input the MFCC features into the dialect speech recognition model, obtain the dialect speech recognition results, and use the dialect speech recognition results as the text labels of its MFCC features and add them to the MFCC feature data set corresponding to the short dialect speech sample; Step 5: At set intervals, the pre-trained convolutional neural network model is retrained using the MFCC feature data set corresponding to the latest short dialect speech sample to obtain the latest dialect speech recognition model.

2. According to claim 1, a method for intelligent recognition of agricultural pests and diseases dialect speech based on deep learning technology is characterized by: The obtaining of the corresponding MFCC features specifically includes: Preprocessing the dialect speech sample / dialect speech data input by the user to obtain a pre-emphasized signal; Perform frame segmentation on the pre-emphasized signal to obtain multiple frame signals, where adjacent frame signals overlap; Perform windowing processing on each frame of signal; Perform fast Fourier transform on each frame of the windowed signal to convert the time domain signal into a frequency domain signal; Design a Mel filter bank, apply the Mel filter bank to each frequency domain signal, and extract the energy of the Mel frequency band; According to the energy of the Mel frequency band, the discrete cosine transform is calculated to obtain the corresponding MFCC features; The corresponding MFCC features are normalized to obtain the corresponding MFCC features.

3. According to claim 2, a method for intelligent recognition of agricultural pests and diseases dialect speech based on deep learning technology is characterized by: The method of preprocessing the dialect speech sample / the dialect speech data input by the user to obtain the pre-emphasized signal specifically includes: The dialect speech sample / the dialect speech data input by the user is pre-emphasized by a first-order high-pass filter, and the first-order high-pass filter is expressed as: , where represents the pre-emphasis coefficient, Represents the signal after pre-emphasis, Represents dialect speech samples / dialect speech data input by the user, express The value of the previous sampling point, Indicates the sampling point number.

4. According to claim 3, a method for intelligent recognition of agricultural pests and diseases dialect speech based on deep learning technology is characterized by: The windowing process for each frame of signal includes: Use the following window function to perform windowing on each frame of signal: ； in, is the length of the window.

5. The method for intelligent recognition of agricultural pests and diseases dialect speech based on deep learning technology according to claim 4 is characterized by: The fast Fourier transform is performed on each frame of the windowed signal to convert the time domain signal into a frequency domain signal, which is expressed as: ； In the formula, represents the window function, represents the index of the frequency component in the frequency domain, is the Fourier basis function, Represents an imaginary unit.

6. The method for intelligent recognition of agricultural pests and diseases dialect speech based on deep learning technology according to claim 5 is characterized by: The Mel scale in the Mel filter bank is expressed as: ； In the formula, Indicates the signal frequency.

7. The method for intelligent recognition of agricultural pests and diseases dialect speech based on deep learning technology according to claim 6 is characterized by: For each frequency domain signal, a Mel filter bank is applied to extract the energy of the Mel frequency band, which is expressed as: ； in, represents the response of the mth Mel filter.

8. The method for intelligent recognition of agricultural pests and diseases dialect speech based on deep learning technology according to claim 7 is characterized by: The discrete cosine transform is calculated based on the energy of the Mel frequency band to obtain the corresponding MFCC features, which is expressed as: ； in, represents MFCC features, and M represents the number of Mel filters in the Mel filter bank.

9. The method for intelligent recognition of agricultural pests and diseases dialect speech based on deep learning technology according to claim 1, characterized in that: The cross entropy loss function is used in the process of training the convolutional neural network model using the long dialect speech sample training set, which is expressed as: ; in, is the true text label of the d-th long dialect speech sample, is the predicted probability of the d-th long dialect speech sample; D is the number of samples in the long dialect speech sample training set; The parameters of the convolutional neural network model are updated by the optimizer, which is expressed as: ; in, is the first-order moment estimate of the gradient, is the second-order moment estimate of the gradient, represents a constant, represents the learning rate, Represents the parameters of the convolutional neural network model in the previous iteration.

10. The method for intelligent recognition of agricultural pests and diseases dialect speech based on deep learning technology according to claim 1, characterized in that: In the process of retraining the pre-trained convolutional neural network model using the short dialect speech sample training set, the optimizer is used to update the model parameters, which is expressed as: ,in, Represents the parameters of the convolutional neural network model of the current round iteration, Represents the loss function About parameters gradient.

Images