Adaptive AI large model streaming speech recognition method, device and equipment

Abstract

The present invention provides a method, device and equipment for streaming speech recognition based on adaptive AI large model, including: performing noise reduction and speech enhancement on the collected streaming speech signal to obtain a preliminary speech signal; performing feature extraction on the preliminary speech signal to obtain a speech feature vector; performing multi-layer perceptron encoding on the speech feature vector to obtain an initial coding vector; dynamically adjusting the clustering center and the number of clusters of the clustering algorithm according to the distribution change of the speech feature vector to cluster the initial coding vector to obtain a clustered coding vector; dynamically assigning weights to the clustered coding vector according to the speech context information through an attention mechanism that adaptively adjusts the weights, and performing weighted processing; inputting the weighted coding vector into a neural network based on an adaptive growth structure for decoding to obtain the speech recognition result. In the present invention, the quality of the speech recognition result is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of speech recognition, and in particular to a method, device and equipment for streaming speech recognition based on an adaptive AI large model. Background Art

[0002] Speech recognition technology has wide application demands in many fields, such as intelligent voice assistants, voice control systems, automatic speech transcription, etc.

[0003] In actual application scenarios, in terms of speech recognition model processing, traditional speech recognition models are mostly designed based on fixed structures and parameters. However, speech data itself is highly diverse and dynamic. Different speakers, accents, language styles, and speech content will cause the distribution of speech features to change. Fixed model structures and parameters cannot adaptively respond to these changes. For example, when faced with new speech patterns or special accents, the model's recognition ability will drop significantly.

[0004] During the neural network decoding stage, the traditional neural network structure cannot adaptively grow and adjust according to the input voice data, and cannot dynamically adapt to different voice feature combinations and changes, resulting in poor quality of speech recognition results. Summary of the invention

[0005] The main purpose of the present invention is to provide an adaptive AI large model streaming speech recognition method, device and equipment, aiming to overcome the defects of poor speech recognition results of current speech recognition models.

[0006] To achieve the above object, the present invention provides a method for streaming speech recognition based on an adaptive AI large model, comprising the following steps:

[0007] Perform noise reduction and speech enhancement on the collected streaming speech signal to obtain a preliminary speech signal;

[0008] Extracting features from the preliminary speech signal to obtain a speech feature vector;

[0009] Inputting the speech feature vector into an adaptive AI large model to obtain a speech recognition result;

[0010] The AI ​​big model is specifically used for:

[0011] Performing multi-layer perceptron encoding on the speech feature vector to obtain an initial encoding vector;

[0012] Dynamically adjusting the clustering center and the number of clusters of the clustering algorithm according to the distribution change of the speech feature vector to cluster the initial coding vector to obtain a clustered coding vector;

[0013] Through the attention mechanism of adaptively adjusting weights, weights are dynamically assigned to the clustered encoding vectors according to the speech context information, and weighted processing is performed;

[0014] The weighted coding vector is input into a neural network based on an adaptive growth structure for decoding to obtain the speech recognition result.

[0015] Furthermore, the weighted coding vector is input into a neural network based on an adaptive growth structure for decoding to obtain the speech recognition result, including:

[0016] Perform feature distribution analysis on the weighted coding vector and dynamically adjust the connection weights between neurons according to the feature distribution;

[0017] Adjust the number of connections between neurons according to the changing characteristics of speech data in time series;

[0018] The weighted coding vector is decoded based on the adjusted neurons to obtain the speech recognition result.

[0019] Furthermore, the weighted coding vector is input into a neural network based on an adaptive growth structure for decoding to obtain the speech recognition result, including:

[0020] The information entropy of the weighted coding vector is detected in real time. When the information entropy is higher than a preset threshold, the number of neurons in the hidden layer of the neural network is increased according to a preset ratio, and the connection weights of the newly added neurons and the surrounding neurons are dynamically initialized according to the correlation of each dimension of the current weighted coding vector; the connection weights between the original neurons are adjusted through a gradient-based optimization algorithm to complete the decoding and obtain the speech recognition result.

[0021] Furthermore, after obtaining the speech recognition result, the method further comprises:

[0022] The speech recognition result is corrected based on the language model and the acoustic model to obtain a streaming speech recognition correction result.

[0023] Furthermore, the performing multi-layer perceptron encoding on the speech feature vector to obtain an initial encoding vector includes:

[0024] Grouping the speech feature vectors according to feature dimensions;

[0025] For each group of speech feature vectors, the intermediate feature vectors are obtained by linear transformation based on the initialization weight matrix by the first layer perceptron;

[0026] Performing nonlinear activation on the intermediate feature vector to obtain an activated feature vector; wherein the activation function dynamically adjusts the slope parameter according to the energy level of the input intermediate feature vector;

[0027] The activated feature vector is input into the second layer perceptron to obtain the initial encoding vector; the weight matrix of the second layer perceptron is optimized through back propagation and mini-batch gradient descent, and the learning rate is dynamically adjusted according to the changes of the activated feature vector in the batch.

[0028] Furthermore, dynamically adjusting the cluster center and the number of clusters of the clustering algorithm according to the distribution change of the speech feature vector to cluster the initial coding vector to obtain the clustered coding vector includes:

[0029] Analyze the probability density distribution of the speech feature vector in each dimension based on the kernel density estimation method to determine the peak value, peak spacing, width and surrounding vector density;

[0030] The cluster center of the clustering algorithm is determined based on the peak value, and the number of clusters is estimated based on the peak spacing, width, and density of surrounding vectors;

[0031] Based on the K-means algorithm, the initial coding vector is assigned to the nearest cluster to obtain a clustered coding vector.

[0032] Furthermore, the attention mechanism of adaptively adjusting weights dynamically assigns weights to the clustered encoding vectors according to the speech context information, and performs weighted processing, including:

[0033] Analyze speech context information, extract semantic key elements, pause features, and intonation trends, and convert them into context feature vectors;

[0034] The correlation between the clustered coding vector and the context feature vector is calculated, the attention weight corresponding to the clustered coding vector is adaptively adjusted according to the correlation, and the attention weight and the clustered coding vector are weightedly calculated.

[0035] The present invention also provides an adaptive AI large model streaming speech recognition device, comprising:

[0036] A processing unit, used to perform noise reduction and voice enhancement on the collected streaming voice signal to obtain a preliminary voice signal;

[0037] An extraction unit, used for performing feature extraction on the preliminary speech signal to obtain a speech feature vector;

[0038] A recognition unit, used for inputting the speech feature vector into an adaptive AI large model to obtain a speech recognition result;

[0039] The AI ​​big model is specifically used for:

[0040] Performing multi-layer perceptron encoding on the speech feature vector to obtain an initial encoding vector;

[0041] Dynamically adjusting the clustering center and the number of clusters of the clustering algorithm according to the distribution change of the speech feature vector to cluster the initial coding vector to obtain a clustered coding vector;

[0042] Through the attention mechanism of adaptively adjusting weights, weights are dynamically assigned to the clustered encoding vectors according to the speech context information, and weighted processing is performed;

[0043] The weighted coding vector is input into a neural network based on an adaptive growth structure for decoding to obtain the speech recognition result.

[0044] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0045] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0046] The method, device and equipment for streaming speech recognition based on an adaptive AI large model provided by the present invention include: performing noise reduction and speech enhancement on the collected streaming speech signal to obtain a preliminary speech signal; performing feature extraction on the preliminary speech signal to obtain a speech feature vector; inputting the speech feature vector into the adaptive AI large model to obtain a speech recognition result; wherein the AI ​​large model is specifically used for: performing multi-layer perceptron encoding on the speech feature vector to obtain an initial encoding vector; dynamically adjusting the clustering center and the number of clusters of the clustering algorithm according to the distribution change of the speech feature vector to cluster the initial encoding vector to obtain a clustered encoding vector; dynamically assigning weights to the clustered encoding vector according to the speech context information through an attention mechanism that adaptively adjusts the weights, and performing weighted processing; inputting the weighted encoding vector into a neural network based on an adaptive growth structure for decoding to obtain the speech recognition result. In the present invention, by decoding in an adaptive AI large model and an adaptive growth structure neural network, the quality of the speech recognition result is improved, and the defect of the poor speech recognition result of the current speech recognition model is overcome. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flowchart of a method for streaming speech recognition based on an adaptive AI large model in one embodiment of the present invention;

[0048] Figure 2 It is a schematic diagram of the steps of a method for streaming speech recognition based on an adaptive AI large model in one embodiment of the present invention;

[0049] Figure 3 It is a structural block diagram of a streaming speech recognition device based on an adaptive AI large model in one embodiment of the present invention;

[0050] Figure 4 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0051] The implementation, functional features and advantages of the present invention will be further described in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0053] Reference Figure 1 as well as Figure 2 In one embodiment of the present invention, a method for streaming speech recognition based on an adaptive AI large model is provided, comprising the following steps:

[0054] Step S1, performing noise reduction and speech enhancement on the collected streaming speech signal to obtain a preliminary speech signal;

[0055] Step S2, extracting features from the preliminary speech signal to obtain a speech feature vector;

[0056] Step S3, inputting the speech feature vector into an adaptive AI large model to obtain a speech recognition result;

[0057] The AI ​​big model is specifically used for:

[0058] Performing multi-layer perceptron encoding on the speech feature vector to obtain an initial encoding vector;

[0059] Dynamically adjusting the clustering center and the number of clusters of the clustering algorithm according to the distribution change of the speech feature vector to cluster the initial coding vector to obtain a clustered coding vector;

[0060] Through the attention mechanism of adaptively adjusting weights, weights are dynamically assigned to the clustered encoding vectors according to the speech context information, and weighted processing is performed;

[0061] The weighted coding vector is input into a neural network based on an adaptive growth structure for decoding to obtain the speech recognition result.

[0062] In this embodiment, an adaptive AI large model streaming speech recognition method is provided, which aims to improve the accuracy and adaptability of speech recognition, and is particularly suitable for processing complex and changeable streaming speech data.

[0063] As described in step S1 above, the collected streaming voice signal is first subjected to noise reduction and voice enhancement processing to obtain a preliminary voice signal. Since the voice signals collected in the actual environment are often interfered by various noises, such as environmental background noise, equipment noise, etc., these noises will seriously affect the accuracy of subsequent voice recognition. Through noise reduction processing, most of the unnecessary noise interference can be removed and the signal-to-noise ratio of the voice signal can be improved. At the same time, voice enhancement can compensate for the attenuation or distortion that may occur during the transmission of voice, making the voice signal clearer and providing higher quality input for subsequent processing steps.

[0064] As described in step S2 above, feature extraction is performed on the preprocessed preliminary speech signal to obtain a speech feature vector. Speech is a complex signal that contains rich information, such as frequency, amplitude, duration, and their changes in time series. Through feature extraction, these complex speech signals can be converted into feature vector forms that are easier to process and analyze. The above feature vector can more comprehensively and accurately reflect the essential characteristics of speech and provide high-quality input data for subsequent model processing.

[0065] As described in the above step S3, speech recognition is performed on the speech feature vector based on the adaptive AI large model to obtain a speech recognition result.

[0066] First, the adaptive AI model performs multi-layer perceptron encoding on the speech feature vector to obtain the initial encoding vector. The multi-layer perceptron can perform nonlinear transformation on the input speech feature vector to extract a higher-level and more abstract feature representation. By reasonably setting the structure and parameters of the multi-layer perceptron, the initial encoding vector can better capture the essence of speech features, laying the foundation for subsequent clustering and processing steps.

[0067] Then, the clustering center and the number of clusters of the clustering algorithm are dynamically adjusted according to the distribution change of the speech feature vector, and the initial coding vector is clustered to obtain the clustered coding vector. Since speech data has a high degree of diversity and dynamic variability, different speakers, accents, language styles and speech contents will cause the distribution of speech features to change. Traditional clustering algorithms usually use fixed clustering centers and cluster numbers, which cannot adapt to such changes. The dynamic clustering method in the present invention can be adjusted according to the real-time distribution of the speech feature vector, so as to more accurately group similar speech features and improve the clustering effect.

[0068] Furthermore, through the attention mechanism of adaptively adjusting weights, weights are dynamically assigned to the clustered coding vectors according to the speech context information, and weighted processing is performed. Speech context information is crucial for understanding speech content. The traditional attention mechanism weight adjustment is not flexible enough and cannot be fully adjusted dynamically according to the speech context information. The attention mechanism in the present invention can adaptively adjust the weight of the coding vector according to the changes in the speech context, so that the feature vector that is more relevant to the current speech context has a greater influence in subsequent processing, thereby improving the accuracy of speech recognition.

[0069] Finally, the weighted coding vector is input into a neural network based on an adaptive growth structure for decoding to obtain a speech recognition result. The traditional neural network structure cannot grow and adjust adaptively according to the input speech data, while the adaptive growth structure neural network in the present invention can dynamically adjust its own structure and parameters according to the input speech data to better adapt to different speech feature combinations and changes. During the decoding process, the neural network can adaptively adjust the connection mode and weight of the neurons according to the characteristics of the weighted coding vector, thereby more accurately decoding the speech recognition result.

[0070] In summary, the adaptive AI large model streaming speech recognition method in this embodiment can effectively improve the accuracy and adaptability of speech recognition and enhance the quality of speech recognition through the above-mentioned innovative processing steps.

[0071] In one embodiment, the step of inputting the weighted coding vector into a neural network based on an adaptive growing structure for decoding to obtain the speech recognition result includes:

[0072] Perform feature distribution analysis on the weighted coding vector and dynamically adjust the connection weights between neurons according to the feature distribution;

[0073] Adjust the number of connections between neurons according to the changing characteristics of speech data in time series;

[0074] The weighted coding vector is decoded based on the adjusted neurons to obtain the speech recognition result.

[0075] In this embodiment, first, the weighted coding vector is subjected to feature distribution analysis. Due to the diversity and complexity of speech data, the weighted coding vector contains different speech feature information, and these features may differ in distribution. By analyzing the feature distribution, the importance, relevance and other information of different features can be understood. According to the analysis results of the feature distribution, the connection weights between neurons are dynamically adjusted. If certain features show high importance in the current speech data, the neuron connection weights associated therewith will be increased accordingly to strengthen the role of these features in the decoding process. On the contrary, if the importance of certain features is low, the connection weights may be reduced. Such dynamic adjustment can make the neural network focus more on key speech features and improve the accuracy and efficiency of decoding.

[0076] Since speech data has the characteristics of change in time series, such as speech speed, intonation, pauses, etc., they will change over time. These changes are crucial for speech recognition, and different time points may require different neural network structures to process. Therefore, according to the changing characteristics of speech data in time series, the number of connections between neurons is adjusted. If at a certain point in time, new complex patterns or changes appear in the speech data, the neural network needs to increase the number of connections to better capture these changes. For example, when there are special accents, multi-language mixtures, etc., increasing the number of connections can give the neural network more paths to process these new features. Conversely, if the speech data is relatively stable over a period of time, the number of connections may be appropriately reduced to reduce computational complexity and increase processing speed.

[0077] Finally, after adjusting the weights and number of connections of the neurons, the structure and parameters of the neural network have adapted to the current characteristics and changes of the speech data. At this time, the weighted coding vector is decoded based on the adjusted neurons, and the coding vector is converted into the final speech recognition result through the calculation and processing process of the neural network. The above result can be the speech content in text form, or other forms of speech information representation. Through this adaptive decoding process, the accuracy and adaptability of speech recognition can be improved to better meet the needs of practical applications.

[0078] In one embodiment, the step of inputting the weighted coding vector into a neural network based on an adaptive growing structure for decoding to obtain the speech recognition result includes:

[0079] The information entropy of the weighted coding vector is detected in real time. When the information entropy is higher than a preset threshold, the number of neurons in the hidden layer of the neural network is increased according to a preset ratio, and the connection weights of the newly added neurons and the surrounding neurons are dynamically initialized according to the correlation of each dimension of the current weighted coding vector; the connection weights between the original neurons are adjusted through a gradient-based optimization algorithm to complete the decoding and obtain the speech recognition result.

[0080] In this embodiment, the weighted coding vector is input into a neural network based on an adaptive growth structure for decoding to obtain a speech recognition result. The key is to dynamically adjust the neural network structure and parameters according to the information entropy of the weighted coding vector and the correlation of each dimension to adapt to different speech data characteristics, improve decoding accuracy and efficiency, and ultimately obtain high-quality speech recognition results.

[0081] First, the information entropy of the weighted coding vector is detected in real time. Information entropy is an indicator of information uncertainty in a system. In speech data, information entropy reflects the complexity and uncertainty of speech signals. By detecting the information entropy of the weighted coding vector in real time, the complexity and changes of the current speech data can be understood. A preset threshold is set in advance to determine whether the complexity of the speech data has reached the level where the neural network structure needs to be adjusted. If the information entropy is higher than the preset threshold, it means that the current speech data is relatively complex and the neural network needs to be adjusted to better process these complex data.

[0082] Furthermore, when the information entropy is higher than a preset threshold, the number of neurons in the hidden layer of the neural network is increased according to a preset ratio. This is because complex speech data requires more computing resources and processing power for accurate decoding. Increasing the number of neurons can expand the capacity of the neural network, enabling it to better capture and process complex speech features. The connection weights of the newly added neurons and the surrounding neurons are dynamically initialized according to the correlation of each dimension of the current weighted coding vector. This ensures that the newly added neurons can work together with the existing neurons and better adapt to the characteristics of the current speech data. By considering the correlation of each dimension of the weighted coding vector, the initialization of the connection weights can be made more reasonable, improving the performance of the neural network.

[0083] The connection weights between the original neurons are adjusted through a gradient-based optimization algorithm. This algorithm can update the connection weights by calculating the gradient based on the error between the output of the neural network and the expected output to minimize the error. In speech recognition, the expected output is usually the correct speech text or other forms of annotated data. Adjusting the connection weights between the original neurons can make the neural network better adapt to the current speech data. As the speech data changes, the original connection weights may no longer be optimal. Adjustment through the optimization algorithm can improve the accuracy and generalization ability of the neural network.

[0084] Finally, after dynamically adjusting the neural network structure and parameters, the neural network can better process the weighted coding vector. Through the decoding process of the neural network, the coding vector is converted into the final speech recognition result. The above recognition result can be the speech content in text form, or other forms of speech information representation. Through this adaptive decoding method, the accuracy and adaptability of speech recognition can be improved to meet the needs of different application scenarios.

[0085] In one embodiment, after obtaining the speech recognition result, the following steps are included:

[0086] The speech recognition result is corrected based on the language model and the acoustic model to obtain a streaming speech recognition correction result.

[0087] In this embodiment, the language model is a statistical model that learns the statistical laws of language based on a large amount of text data, including the frequency of occurrence of words, the collocation relationship between words, grammatical structure, etc. In speech recognition, the language model can judge the rationality of the speech recognition result based on the context information and statistical laws of the language.

[0088] The preliminary speech recognition results are input into the language model. The language model will analyze the probability of occurrence of each word in the recognition results and the rationality of the collocation between words based on the learned language rules. If it is found that the probability of occurrence of certain words is low or the collocation between words does not conform to common language habits, the language model will make possible correction suggestions. For example, if an uncommon word appears in the recognition result, and according to the statistics of the language model, another common word has a higher probability of occurrence in the same context, then the language model may suggest replacing the uncommon word with a common word.

[0089] The above acoustic model is mainly used to describe the relationship between speech signals and speech features. It learns the speech pronunciation patterns corresponding to different speech features by training a large amount of speech data. In speech recognition, the acoustic model can judge the acoustic rationality of the speech recognition results based on the input speech signal features.

[0090] The speech features corresponding to the preliminary speech recognition results are compared with the acoustic model. If it is found that some parts of the speech features are inconsistent with the predictions of the acoustic model, the acoustic model will make possible correction suggestions. For example, if the pronunciation of a syllable deviates from the common pronunciation pattern in terms of acoustic features, the acoustic model may suggest adjusting the recognition result of the syllable to make it more in line with the acoustic rules.

[0091] In one embodiment, the step of performing multi-layer perceptron encoding on the speech feature vector to obtain an initial encoding vector includes:

[0092] Grouping the speech feature vectors according to feature dimensions;

[0093] For each group of speech feature vectors, the intermediate feature vectors are obtained by linear transformation based on the initialization weight matrix by the first layer perceptron;

[0094] Performing nonlinear activation on the intermediate feature vector to obtain an activated feature vector; wherein the activation function dynamically adjusts the slope parameter according to the energy level of the input intermediate feature vector;

[0095] The activated feature vector is input into the second layer perceptron to obtain the initial encoding vector; the weight matrix of the second layer perceptron is optimized through back propagation and mini-batch gradient descent, and the learning rate is dynamically adjusted according to the changes of the activated feature vector in the batch.

[0096] In this embodiment, the speech feature vector is mainly encoded by a multi-layer perceptron to obtain an initial encoding vector.

[0097] First, the speech feature vectors are grouped according to the feature dimensions. The speech feature vectors usually contain information of multiple dimensions, such as frequency, amplitude, duration and other different speech attributes. Grouping by feature dimensions can group feature vectors with similar attributes into one group, which is convenient for subsequent targeted processing of different feature attributes. This grouping method can better understand and utilize speech features of different dimensions and improve the accuracy and effectiveness of encoding.

[0098] Then, for each group of speech feature vectors, the first layer of perceptron performs linear transformation based on the initialization weight matrix to obtain the intermediate feature vector. The first layer of perceptron performs linear transformation on each group of speech feature vectors. Linear transformation can map the original feature vector to a new feature space, thereby extracting some potential and more representative features.

[0099] The weight matrix is ​​initialized based on the distribution statistics of the corresponding features in a large amount of pre-trained speech data. This allows the first layer of perceptrons to perform preliminary feature extraction based on existing knowledge and experience when processing new speech feature vectors, thereby improving the efficiency and accuracy of encoding.

[0100] Then, the intermediate feature vector is nonlinearly activated to obtain the activated feature vector. The introduction of nonlinear activation function is to enable the model to learn more complex nonlinear relationships. Speech data often has complex nonlinear features, and simple linear transformation cannot fully capture these features. Through nonlinear activation, the expressive power of the model can be increased and the nonlinear information in the speech features can be better extracted. The activation function dynamically adjusts the slope parameter according to the energy level of the input intermediate feature vector. This makes the activation function more adaptable to different input data and improves the adaptability of the model to different speech features. For example, when the energy level of the intermediate feature vector is high, the slope parameter of the activation function can be appropriately increased to enhance the response to strong signals; when the energy level is low, the slope parameter can be reduced to avoid over-activation.

[0101] Finally, the activated feature vector is input into the second layer perceptron to obtain the initial encoding vector. The second layer perceptron further processes the feature vector after nonlinear activation. It can learn higher-level feature representation, integrate feature information of different groups, and provide richer and more accurate initial encoding vectors for the final speech recognition. The weight matrix of the second layer perceptron is optimized by back propagation and mini-batch gradient descent. The back propagation algorithm can back propagate the error signal according to the error between the output of the model and the expected output, and update the weight matrix to minimize the error. Mini-batch gradient descent divides the training data into small batches for processing, and only uses a small batch of data each time the weight matrix is ​​updated, which can improve the efficiency and stability of training. The learning rate is dynamically adjusted according to the change of the activated feature vector in the batch. The learning rate is a parameter that controls the amplitude of weight update. Dynamically adjusting the learning rate can make the model better adapt to different data distributions and changes during the training process. For example, when the activated feature vector changes greatly in the batch, the learning rate can be appropriately reduced to avoid the instability of the model caused by too fast weight update; when the change is small, the learning rate can be appropriately increased to speed up the convergence of the model.

[0102] In one embodiment, dynamically adjusting the cluster center and the number of clusters of the clustering algorithm according to the distribution change of the speech feature vector to cluster the initial coding vector to obtain the clustered coding vector includes:

[0103] Analyze the probability density distribution of the speech feature vector in each dimension based on the kernel density estimation method to determine the peak value, peak spacing, width and surrounding vector density;

[0104] The cluster center of the clustering algorithm is determined based on the peak value, and the number of clusters is estimated based on the peak spacing, width, and density of surrounding vectors;

[0105] Based on the K-means algorithm, the initial coding vector is assigned to the nearest cluster to obtain a clustered coding vector.

[0106] In this embodiment, first, the probability density distribution of the speech feature vector in each dimension is analyzed based on the kernel density estimation method. Kernel density estimation is a non-parametric estimation method used to estimate the probability density function of a random variable. In this embodiment, the probability density distribution of the speech feature vector in each dimension can be analyzed by the kernel density estimation method. It is helpful to understand the distribution characteristics of the speech feature vector in different dimensions, including information such as peak values, valley values, and the width of the distribution. The peak value, peak spacing, width, and density of the surrounding vectors of the speech feature vector in each dimension can be determined by the kernel density estimation method. The peak value represents an area with a higher probability density, corresponding to some representative speech feature patterns. The peak spacing, width, and density of the surrounding vectors can reflect the relationship between different peak values ​​and the distribution of the feature vectors in each area.

[0107] Then, the cluster center is determined based on the peak value, and the number of clusters is estimated based on the peak distance, etc. The cluster center of the clustering algorithm is determined based on the peak value. Usually, the peak position or the value near it is used as the potential cluster center. This can make the cluster center closer to the actual distribution of the data and improve the accuracy of clustering.

[0108] The number of clusters is estimated based on the peak spacing, width, and density of surrounding vectors. If the peaks are close together and the surrounding vectors are densely distributed, it means that the speech features in these areas are similar, and the number of clusters can be appropriately reduced; if the peaks are far apart and the surrounding vectors are sparsely distributed, it indicates that there are different speech feature patterns, and the number of clusters needs to be increased. In this way, the number of clusters can be dynamically adjusted according to the actual distribution of the data, improving the adaptability of clustering.

[0109] Finally, clustering is performed based on the K-means algorithm. The K-means algorithm is a commonly used clustering algorithm that assigns data points to the nearest cluster center and continuously updates the position of the cluster center until the convergence condition is reached. In this technical solution, the K-means algorithm is used to cluster the initial coding vectors. For each initial coding vector, its distance from each cluster center is calculated (the distance measurement method can be selected according to the properties of the speech features, such as Euclidean distance, cosine distance, etc.), and it is assigned to the cluster with the nearest distance. In this way, similar initial coding vectors can be classified into the same category to obtain clustered coding vectors.

[0110] In summary, this technical solution can effectively group the initial coding vectors by analyzing the distribution of speech feature vectors, dynamically adjusting the parameters of the clustering algorithm, and using the K-means algorithm for clustering, thereby providing more valuable information for subsequent speech recognition processing.

[0111] In one embodiment, the attention mechanism of adaptively adjusting weights dynamically assigns weights to the clustered encoding vectors according to the speech context information and performs weighted processing, including:

[0112] Analyze speech context information, extract semantic key elements, pause features, and intonation trends, and convert them into context feature vectors;

[0113] The correlation between the clustered coding vector and the context feature vector is calculated, the attention weight corresponding to the clustered coding vector is adaptively adjusted according to the correlation, and the attention weight and the clustered coding vector are weightedly calculated.

[0114] In this embodiment, first, the speech context information is analyzed and converted into a context feature vector. Speech context information is crucial for accurately understanding the speech content. This includes the speech information before and after the current speech segment in the speech stream, covering multiple levels such as semantics, grammar, and pragmatics. Specifically, semantic key elements are extracted, that is, words or phrases that play an important role in the meaning of the speech are identified. For example, in a sentence, core words such as nouns and verbs are often semantic key elements. At the same time, attention is paid to pause features. The position and duration of the pause can imply the structure and rhythm of the sentence, which helps to understand the segmentation and focus of semantics. In addition, the intonation trend is analyzed, such as rising tone, falling tone, flat tone, etc. Different intonations can convey different emotions and tones, which also affect the understanding of the meaning of the speech.

[0115] Then, the extracted semantic key elements, pause features, intonation trends and other information are converted into quantifiable context feature vectors through specific methods. For semantic key elements, the word vector representation method can be used to map each keyword into a high-dimensional vector, which can capture the semantic information of the vocabulary. For pause features, numerical values ​​can be used to represent the duration of the pause or the proportion of the pause position in the speech stream. For intonation trends, audio features, such as changes in fundamental frequency, can be extracted and converted into numerical feature vectors.

[0116] Then, the correlation is calculated and the weights are adjusted adaptively for weighted processing, and the correlation between the clustered coding vector and the context feature vector is calculated. This correlation measures the degree of association between the speech features represented by the clustered coding vector and the current speech context. A specific correlation measurement function can be pre-designed to comprehensively consider the similarity between the speech features represented by the coding vector and the context feature vector in terms of semantics, speech rhythm, etc. For example, if the speech features represented by the coding vector are highly correlated with a topic in the context semantics, or its rhythm matches the pause feature, then the correlation value will be high.

[0117] The attention weights corresponding to the clustered coding vectors are adaptively adjusted according to the correlation calculation results. If the correlation is high, it means that the coding vector is more relevant to the current speech context, so the corresponding attention weight is increased. The magnitude of the weight adjustment is proportional to the strength of the correlation, that is, the higher the correlation, the greater the increase in weight. At the same time, in order to avoid excessive concentration of weights, an upper limit of the weight is set to ensure that the weights of other coding vectors are not overly compressed due to the high correlation between a certain coding vector and the context.

[0118] Finally, the adjusted attention weights are weighted with the clustered encoding vectors. In the subsequent processing, the weighted encoding vectors that are more relevant to the speech context information will have a greater influence. This can make the speech recognition process pay more attention to the speech features related to the current context, improving the accuracy and adaptability of speech recognition.

[0119] In one embodiment, after obtaining the speech recognition result, the following steps are included:

[0120] Extracting text features from the speech recognition result to obtain first feature information; the first feature information includes text length, number of characters, and number of text types;

[0121] Obtaining version identification information of the adaptive AI large model; discretely distributing the first feature information into the version identification information to obtain a version string;

[0122] Adding the characters in the version string to each node of a preset binary tree template one by one in sequence to obtain a character binary tree; adding the characters in the version string to a preset data table one by one in sequence to obtain a character data table;

[0123] Extracting features from the character binary tree and the character data table respectively to obtain corresponding binary tree features and data table features;

[0124] A feature key is generated based on the binary tree feature and the data table feature; and the speech recognition result is encrypted based on the feature key.

[0125] In this embodiment, first, text feature extraction is performed on the speech recognition result. In order to further process the speech recognition result, text feature extraction is first performed to obtain key information about the recognition result for subsequent encryption and version management operations. Among them, the first feature information includes text length, number of characters, and number of text types; the text length is the total length of all characters in the speech recognition result, which can reflect the size of the recognition result. The number of characters indicates the total number of characters contained in the recognition result, which helps to understand the richness of the content. The number of text types is the number of different types of characters, such as the number of types of Chinese characters, numbers, letters, special symbols, etc., which can provide information about the diversity of recognition results.

[0126] Then, obtain the version identification information of the adaptive AI big model. The version identification information can clearly indicate the specific version of the AI ​​big model used. Different versions may differ in performance, function or algorithm. Obtaining the version identification information will help to process and manage the speech recognition results more targetedly.

[0127] Then, the first feature information is discretely distributed to the version identification information, and the previously extracted text feature information (text length, number of characters, and number of text types) is distributed to the version identification information in a preset manner to obtain a version string. This process can be an encoding operation, which combines the text features with the version information to form a string with a specific structure for subsequent processing.

[0128] Then, the characters in the version string are added one by one to each node of the preset binary tree template to obtain a character binary tree. A binary tree is a data structure with efficient search, insertion and deletion performance. By adding the characters of the version string to the binary tree, a binary tree structure that reflects the version string structure and character distribution can be constructed. Each node can store a character, and the position in the binary tree is determined according to the size relationship of the characters, thereby forming a hierarchical structure. At the same time, the characters in the version string are added one by one to a preset data table to obtain a character data table. The data table can be a simple table structure, and each row can store a character and related information, such as the position of the character in the version string, the number of occurrences, etc.

[0129] Furthermore, feature extraction is performed on the constructed binary tree of characters, and the extracted features include the height of the binary tree, the number of nodes, the distribution of leaf nodes, etc. The above features can reflect the distribution and complexity of the version string in the binary tree structure.

[0130] Feature extraction is performed on the character data table, which may include features such as frequency distribution, position information, and continuous appearance patterns of characters. The above features can provide statistical and structural information about the version string in the data table.

[0131] Finally, based on the extracted binary tree features and data table features, a feature key is generated through a preset algorithm. The above key can be a numerical value, a string or other forms of data. It is generated based on the features of the version string and is unique and specific. The speech recognition result is encrypted using the generated feature key. Encryption can use common encryption algorithms, such as symmetric encryption or asymmetric encryption, to convert the speech recognition result into ciphertext to protect its security and privacy. Only users with the correct decryption key can restore the ciphertext to the original speech recognition result.

[0132] In one embodiment, based on the binary tree feature and the data table feature, a feature key is generated, specifically including:

[0133] For binary tree features, the height, number of nodes, number of leaf nodes, and distribution ratio of nodes at different levels of the binary tree are first quantified. For example, the height of the binary tree is multiplied by a preset first coefficient, the number of nodes is multiplied by a second coefficient, and the number of leaf nodes is multiplied by a third coefficient. The distribution ratio of nodes at different levels is converted into a vector form through a specific mapping function. The above coefficients and mapping functions can be determined based on experimental data and cryptographic security requirements, aiming to give different features appropriate weights and representations in subsequent combinations.

[0134] For data table features, the frequency distribution of characters is quantified by histogram statistics, and the frequency value of each character is multiplied by a fourth coefficient related to the character type; the position information is converted into a normalized value related to the data table size through relative position encoding, and then multiplied by the fifth coefficient; the continuously occurring patterns are converted into pattern encoding vectors through pattern recognition algorithms, and each dimension value of the vector is multiplied by the sixth coefficient. These coefficients are also determined based on experiments and security requirements.

[0135] Then, the quantized binary tree features and data table features are fused. A fusion method based on hash chain is adopted. The binary tree feature vector is first used as the initial input of the hash chain, and then the elements in the data table feature vector are added to the calculation of the hash chain one by one in a specific order. The hash function selects a cryptographic hash function with high security and low collision rate, and its parameters are configured according to the overall security level of the speech recognition system.

[0136] After the hash chain calculation is completed, the obtained hash value is further processed. The specific byte segment of the hash value is extracted, and a pseudo-random number sequence is generated according to the value of the byte segment. The generation algorithm of the sequence is based on a cryptographically secure pseudo-random number generator and initialized with the version identification information of the speech recognition system to ensure that different sequences are generated in different version systems.

[0137] Finally, the pseudo-random number sequence is used to generate a feature key through a custom key construction algorithm. The algorithm adjusts and combines the pseudo-random number sequence according to the language type of the speech recognition result, the dimension of the speech feature vector and other information, so that the feature key is closely related to the speech recognition process and has high confidentiality and uniqueness.

[0138] In another embodiment, a feature key is generated based on the binary tree feature and the data table feature, specifically including:

[0139] For binary tree features, the height of the binary tree, the number of nodes, the number of leaf nodes, the information entropy of nodes at each layer, and the complexity of the subtree structure are quantified. Specifically, the height value of the binary tree is converted into a vector in a high-dimensional space through a nonlinear transformation function (such as a combination of logarithmic functions and trigonometric functions), and the number of nodes and the number of leaf nodes are multiplied by the complexity parameter of the current speech recognition task (dynamically determined according to the dimension of the speech feature vector, speech duration, etc.) and then subjected to exponential operation to obtain the quantified value. The information entropy of nodes at each layer is obtained by weighting the semantic relevance of the speech context through the improved Shannon entropy calculation method, and the complexity of the subtree structure is quantified by associating the structural measurement algorithm in graph theory with the speech spectrum features. The above quantized values ​​are dynamically mapped, and the mapping function is adjusted in real time according to the language type of the speech recognition result (such as different mapping parameters for different languages ​​such as Chinese and English) and the adaptive adjustment parameters of the current AI large model (such as clustering center, weight, etc.), so that each feature has a unique quantitative mapping under different languages ​​and model states.

[0140] For data table features, the character frequency distribution is quantified using the fractal dimension calculation method and converted into a fractal dimension vector. Each dimension represents the fractal characteristics of different frequency intervals. This method can capture the complex distribution pattern of character frequencies. Position information is mapped into a hyperbolic space to convert linear position data into coordinates in the hyperbolic space. The curvature parameters of the hyperbolic space are dynamically adjusted according to the real-time accuracy of speech recognition to highlight key position information. Continuously occurring patterns are extracted through autoencoders in deep learning to compress the patterns into low-dimensional semantic vectors. The training data of the autoencoder comes from a large-scale speech text corpus and is fine-tuned according to the current text style of speech recognition (such as formal, colloquial, etc.) to make the pattern vector more targeted.

[0141] A feature fusion method based on quantum entanglement is adopted to treat the quantized binary tree features and data table features as two quantum states. By simulating the quantum entanglement phenomenon, a quantum entanglement matrix is ​​constructed, whose element calculation involves the cross terms of the binary tree and data table features, and the value of each element is dynamically updated according to the quantum random number generator (generating true random numbers based on quantum physics principles) and the acoustic model parameters of speech recognition (such as spectral features, resonance peak information, etc.), realizing the deep fusion of cross-modal features at the quantum level.

[0142] The fused quantum state is measured, and the selection of the measurement basis is based on the rhythmic features of the speech (such as intonation, stress, etc.) and the dynamic distribution changes of the speech feature vector. The measurement results are passed through a bio-inspired key generation algorithm, which simulates the signal transmission and encryption mechanism in the biological nervous system, such as borrowing the pulse coding and synaptic plasticity principles of neurons, to convert the measurement results into the initial key sequence.

[0143] Finally, the initial key sequence is adjusted using the semantic grammatical structure in the speech recognition results. Based on the semantic role annotations and grammatical dependencies, the elements in the key sequence are permuted, replaced, and extended, and the final feature key is generated by combining the hash value in the version identification information of the current adaptive AI large model (calculated using a hash algorithm that is resistant to quantum attacks). This feature key is tightly coupled with multiple levels of the speech recognition process, has extremely high creativity and security, can effectively resist various known and unknown attack methods, and protect the privacy and security of speech recognition results.

[0144] Reference Figure 3 In another embodiment of the present invention, there is provided a streaming speech recognition device based on an adaptive AI large model, comprising:

[0145] A processing unit, used to perform noise reduction and voice enhancement on the collected streaming voice signal to obtain a preliminary voice signal;

[0146] An extraction unit, used for performing feature extraction on the preliminary speech signal to obtain a speech feature vector;

[0147] A recognition unit, used for inputting the speech feature vector into an adaptive AI large model to obtain a speech recognition result;

[0148] The AI ​​big model is specifically used for:

[0149] Performing multi-layer perceptron encoding on the speech feature vector to obtain an initial encoding vector;

[0150] Dynamically adjusting the clustering center and the number of clusters of the clustering algorithm according to the distribution change of the speech feature vector to cluster the initial coding vector to obtain a clustered coding vector;

[0151] Through the attention mechanism of adaptively adjusting weights, weights are dynamically assigned to the clustered encoding vectors according to the speech context information, and weighted processing is performed;

[0152] The weighted coding vector is input into a neural network based on an adaptive growth structure for decoding to obtain the speech recognition result.

[0153] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0154] Reference Figure 4 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 4 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0155] Those skilled in the art will understand that Figure 4 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0156] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0157] In summary, the method, device and equipment for streaming speech recognition based on an adaptive AI large model provided in the embodiment of the present invention include: denoising and enhancing the collected streaming speech signal to obtain a preliminary speech signal; extracting features from the preliminary speech signal to obtain a speech feature vector; inputting the speech feature vector into the adaptive AI large model to obtain a speech recognition result; wherein the AI ​​large model is specifically used to: perform multi-layer perceptron encoding on the speech feature vector to obtain an initial coding vector; dynamically adjust the clustering center and the number of clusters of the clustering algorithm according to the distribution change of the speech feature vector to cluster the initial coding vector to obtain a clustered coding vector; dynamically assign weights to the clustered coding vector according to the speech context information through the attention mechanism of adaptively adjusting the weights, and perform weighted processing; input the weighted coding vector into a neural network based on an adaptive growth structure for decoding to obtain the speech recognition result. In the present invention, by decoding in the adaptive AI large model and the neural network of the adaptive growth structure, the quality of the speech recognition result is improved, and the defect of the poor speech recognition result of the current speech recognition model is overcome.

[0158] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0159] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0160] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A streaming speech recognition method based on an adaptive AI large model, characterized in that: The following steps are involved: Perform noise reduction and speech enhancement on the collected streaming speech signal to obtain a preliminary speech signal; Extracting features from the preliminary speech signal to obtain a speech feature vector; Inputting the speech feature vector into an adaptive AI large model to obtain a speech recognition result; The AI ​​big model is specifically used for: Performing multi-layer perceptron encoding on the speech feature vector to obtain an initial encoding vector; Dynamically adjusting the clustering center and the number of clusters of the clustering algorithm according to the distribution change of the speech feature vector to cluster the initial coding vector to obtain a clustered coding vector; Through the attention mechanism of adaptively adjusting weights, weights are dynamically assigned to the clustered encoding vectors according to the speech context information, and weighted processing is performed; Inputting the weighted coding vector into a neural network based on an adaptive growth structure for decoding to obtain the speech recognition result; The step of performing multi-layer perceptron encoding on the speech feature vector to obtain an initial encoding vector includes: Grouping the speech feature vectors according to feature dimensions; For each group of speech feature vectors, the intermediate feature vectors are obtained by linear transformation based on the initialization weight matrix by the first layer perceptron; Performing nonlinear activation on the intermediate feature vector to obtain an activated feature vector; wherein the activation function dynamically adjusts the slope parameter according to the energy level of the input intermediate feature vector; The activation feature vector is input into the second layer perceptron to obtain the initial encoding vector; the weight matrix of the second layer perceptron is optimized through back propagation and mini-batch gradient descent, and the learning rate is dynamically adjusted according to the changes of the activation feature vector in the batch; Dynamically adjusting the cluster center and the number of clusters of the clustering algorithm according to the distribution change of the speech feature vector to cluster the initial coding vector to obtain the clustered coding vector, including: Analyze the probability density distribution of the speech feature vector in each dimension based on the kernel density estimation method to determine the peak value, peak spacing, width and surrounding vector density; The cluster center of the clustering algorithm is determined based on the peak value, and the number of clusters is estimated based on the peak spacing, width, and density of surrounding vectors; Based on the K-means algorithm, the initial coding vector is assigned to the nearest cluster to obtain a clustered coding vector.

2. The method for streaming speech recognition based on an adaptive AI large model according to claim 1, characterized in that: The step of inputting the weighted coding vector into a neural network based on an adaptive growth structure for decoding to obtain the speech recognition result includes: Perform feature distribution analysis on the weighted coding vector and dynamically adjust the connection weights between neurons according to the feature distribution; Adjust the number of connections between neurons according to the changing characteristics of speech data in time series; The weighted coding vector is decoded based on the adjusted neurons to obtain the speech recognition result.

3. The method for streaming speech recognition based on an adaptive AI large model according to claim 1, characterized in that: The step of inputting the weighted coding vector into a neural network based on an adaptive growth structure for decoding to obtain the speech recognition result includes: The information entropy of the weighted coding vector is detected in real time. When the information entropy is higher than a preset threshold, the number of neurons in the hidden layer of the neural network is increased according to a preset ratio, and the connection weights of the newly added neurons and the surrounding neurons are dynamically initialized according to the correlation of each dimension of the current weighted coding vector; the connection weights between the original neurons are adjusted through a gradient-based optimization algorithm to complete the decoding and obtain the speech recognition result.

4. The adaptive AI large model streaming speech recognition method according to claim 1 is characterized in that: After the speech recognition result is obtained, the method includes: The speech recognition result is corrected based on the language model and the acoustic model to obtain a streaming speech recognition correction result.

5. The adaptive AI large model streaming speech recognition method according to claim 1 is characterized in that: The attention mechanism of adaptively adjusting weights dynamically assigns weights to the clustered encoding vectors according to the speech context information and performs weighted processing, including: Analyze speech context information, extract semantic key elements, pause features, and intonation trends, and convert them into context feature vectors; The correlation between the clustered coding vector and the context feature vector is calculated, the attention weight corresponding to the clustered coding vector is adaptively adjusted according to the correlation, and the attention weight and the clustered coding vector are weightedly calculated.

6. A streaming speech recognition device based on an adaptive AI large model, characterized in that: include: A processing unit, used to perform noise reduction and voice enhancement on the collected streaming voice signal to obtain a preliminary voice signal; An extraction unit, used for performing feature extraction on the preliminary speech signal to obtain a speech feature vector; A recognition unit, used for inputting the speech feature vector into an adaptive AI large model to obtain a speech recognition result; The AI ​​big model is specifically used for: Performing multi-layer perceptron encoding on the speech feature vector to obtain an initial encoding vector; Dynamically adjusting the clustering center and the number of clusters of the clustering algorithm according to the distribution change of the speech feature vector to cluster the initial coding vector to obtain a clustered coding vector; Through the attention mechanism of adaptively adjusting weights, weights are dynamically assigned to the clustered encoding vectors according to the speech context information, and weighted processing is performed; Inputting the weighted coding vector into a neural network based on an adaptive growth structure for decoding to obtain the speech recognition result; The step of performing multi-layer perceptron encoding on the speech feature vector to obtain an initial encoding vector includes: Grouping the speech feature vectors according to feature dimensions; For each group of speech feature vectors, the intermediate feature vectors are obtained by linear transformation based on the initialization weight matrix by the first layer perceptron; Performing nonlinear activation on the intermediate feature vector to obtain an activated feature vector; wherein the activation function dynamically adjusts the slope parameter according to the energy level of the input intermediate feature vector; The activation feature vector is input into the second layer perceptron to obtain the initial encoding vector; the weight matrix of the second layer perceptron is optimized through back propagation and mini-batch gradient descent, and the learning rate is dynamically adjusted according to the changes of the activation feature vector in the batch; Dynamically adjusting the cluster center and the number of clusters of the clustering algorithm according to the distribution change of the speech feature vector to cluster the initial coding vector to obtain the clustered coding vector, including: Analyze the probability density distribution of the speech feature vector in each dimension based on the kernel density estimation method to determine the peak value, peak spacing, width and surrounding vector density; The cluster center of the clustering algorithm is determined based on the peak value, and the number of clusters is estimated based on the peak spacing, width, and density of surrounding vectors; Based on the K-means algorithm, the initial coding vector is assigned to the nearest cluster to obtain a clustered coding vector.

7. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images