Hainan dialect speech recognition optimization system based on knowledge distillation

Abstract

The application discloses a Hainan dialect speech recognition optimization system based on knowledge distillation, comprising: a data preprocessing module for processing Hainan dialect speech data and Hainan dialect text data, outputting MFCC speech feature sequences, labeled text labels and Hainan dialect text data; a teacher model module comprising an RNN language model, a CNN language model and a Transformer language model, wherein the dialect text data is input into the RNN language model, the CNN language model and the Transformer language model, and soft labels and intermediate layer features are obtained through dynamic temperature adjustment; a student model training module for knowledge distillation and parameter optimization of the student model according to the MFCC speech feature sequences, the labeled text labels, the soft labels and the intermediate layer features; and an output module for speech recognition of the Hainan dialect by using the student model after knowledge distillation and parameter optimization to obtain a recognition result; and the trained student model is small in size and low in calculation complexity, and finally an accurate recognition result can be obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of speech recognition, in particular to a Hainan dialect speech recognition optimization system based on knowledge distillation. BACKGROUND

[0002] The protection of dialects and the development of intelligent recognition technology have become the focus of social attention. At present, although the Hainan dialect speech recognition technology has achieved certain results, its recognition accuracy still needs to be improved in actual application. The primary constraint factor is the limited training data scale and the insufficient model generalization ability. Although research teams are working hard to build a Hainan dialect speech database, compared with languages such as Mandarin and Cantonese that have established mature corpus systems, the construction of Hainan dialect speech resource library presents obvious lag, which restricts the training and performance optimization of the model. At the same time, the adaptability of the technology in actual application scenarios is slightly insufficient. In the face of complex environmental noise, various speaker styles and diverse language expression methods, the response ability of the Hainan dialect speech recognition system still needs to be improved, which also limits its wide application in tourism, medical care, and the elderly care field.

[0003] Compared with the scarce Hainan dialect audio-text paired data, the cost of obtaining pure text data in actual production is indeed lower, and the number of available pure text data is often several or even tens of orders of magnitude more than the audio-text paired data. It is very important to integrate an external language model into a speech recognition system. Although the traditional methods such as shallow fusion or deep fusion are effective, the model is relatively complex, and the calculation amount is large during decoding. SUMMARY

[0004] In view of the above prior art, the present application provides a Hainan dialect speech recognition optimization system based on knowledge distillation, which mainly solves the technical problems existing in the background art.

[0005] To achieve the above purpose, the technical scheme of the embodiment of the present application is as follows:

[0006] The Hainan dialect speech recognition optimization system based on knowledge distillation comprises:

[0007] A data preprocessing module for processing Hainan dialect speech data and Hainan dialect text data, outputting MFCC speech feature sequences, labeled text labels, and Hainan dialect text data;

[0008] A teacher model module comprising an RNN language model, a CNN language model, and a Transformer language model, wherein the dialect text data is input into the RNN language model, the CNN language model, and the Transformer language model, respectively, and soft labels and intermediate layer features are obtained after dynamic temperature adjustment;

[0009] a student model training module configured to perform knowledge distillation and parameter tuning on the student model according to the MFCC speech feature sequence, the annotated text label, the soft label, and the intermediate layer feature;

[0010] an output module configured to perform speech recognition on the Hainan dialect using the student model after knowledge distillation and parameter tuning to obtain a recognition result.

[0011] Optionally, the RNN language model is an LSTM-based recurrent neural network language model configured to capture sequence dependency in the Hainan dialect text data.

[0012] The CNN language model is a convolutional neural network-based language model configured to capture local features of the Hainan dialect text data.

[0013] The Transformer language model is stacked with multiple Transformer encoders and configured to extract multi-scale features of the Hainan dialect text data.

[0014] Optionally, the knowledge distillation and parameter tuning on the student model according to the MFCC speech feature sequence, the annotated text label, the soft label, and the intermediate layer feature include:

[0015] For the MFCC speech feature sequence, a prediction result is generated by forward propagation using the student model.

[0016] For the annotated text label, the annotated text label is taken as a true label, a cross-entropy loss between the prediction result and the true label is calculated, and a classification loss is obtained.

[0017] For the soft label, the soft labels obtained by the RNN language model, the CNN language model, and the Transformer language model are weighted and averaged, a distillation loss between the prediction result of the student model and the weighted and averaged soft label is calculated.

[0018] For the intermediate layer feature, real-time intermediate layer features corresponding to the teacher model in the student model are extracted, the intermediate layer features output by the teacher model are matched with the real-time intermediate layer features extracted by the student model, a cosine similarity loss between the intermediate layer features of the teacher model and the real-time intermediate layer features of the student model is calculated, and an intermediate layer feature matching loss is obtained.

[0019] A total loss function is calculated based on the classification loss, the distillation loss, and the intermediate layer feature matching loss, the student model is trained according to the total loss function, and a trained student model is obtained.

[0020] Optionally, the calculation formula of the classification loss is:

[0021]

[0022] wherein y is a real label, t is a temperature, q (t) is a probability distribution output by a student model after softening; is a classification loss.

[0023] Optionally, the calculation formula of the distillation loss is:

[0024]

[0025] wherein, is a distillation loss, t is a temperature, p (t) is a probability distribution output by a teacher model after softening, M is a number of teacher models, is a soft label of the mth teacher model after temperature processing, w m is a weight;

[0026] The weight calculation formula in the distillation loss is:

[0027]

[0028] Confidence m =1 / CER m

[0029] wherein Confidence m is a confidence of the mth teacher model.

[0030] Optionally, the calculation formula of the intermediate layer feature matching loss is:

[0031]

[0032] wherein L is a number of intermediate layers, is a feature of the lth layer of a teacher model, is a feature of the lth layer of a student model.

[0033] Optionally, the expression of the total loss function is:

[0034]

[0035] wherein a and β are loss weights, is a total loss, is a classification loss, is a distillation loss, is an intermediate layer feature matching loss.

[0036] The beneficial effects of the present application are that the Hainan dialect speech recognition optimization system based on knowledge distillation provided by the present application pre-processes Hainan dialect speech data and Hainan dialect text data through a data pre-processing module, extracts an MFCC speech feature sequence, obtains text labels and Hainan dialect text data, processes the Hainan dialect text data processed by the data pre-processing module using a teacher model module, the teacher model module includes multiple language models, the multiple language models are fused to generate soft labels and intermediate layer features, and then a student model training module is used for knowledge distillation and parameter optimization, classification loss, distillation loss and intermediate layer feature matching loss are calculated, a total loss function is obtained, the training process is optimized through the total loss function, and finally a trained student model is obtained, the Hainan dialect is recognized using the trained student model, and a speech recognition result is obtained; in the training stage, the training of the student model is guided by generating soft labels using multiple external language models, and in the decoding stage, no additional external language model needs to be introduced, thereby reducing the computational complexity; and the knowledge of the teacher model is migrated to the student model, that is, the knowledge in the complex model is migrated to the simple model, the parameters are fewer, and the computational complexity is lower. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The structure schematic diagram of the Hainan dialect speech recognition optimization system based on knowledge distillation provided in the embodiments of the present application is shown.

[0038] Figure 2 The loss function construction schematic diagram provided in the embodiments of the present application is shown. DETAILED DESCRIPTION

[0039] The technical solutions of the present application are further described in detail below in combination with the drawings and specific embodiments. Unless otherwise defined, all technical and scientific terms used herein have the same meanings as understood by those skilled in the art to which the present application belongs. The terms used in the specification of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. In the following description, the expression "some embodiments" describes a subset of all possible embodiments, but it should be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.

[0040] In the following description, a large number of specific details are given in order to provide a more thorough understanding of the present application. However, it is obvious to those skilled in the art that the present application can be implemented without one or more of these details. In other examples, some technical features known in the art are not described in order to avoid obscuring the present application.

[0041] It is to be understood that the application can assume various alternative orientations, except where expressly omitted. Accordingly, these embodiments are not intended to limit the scope of the application as described herein. Rather, the scope of the application is limited only by the appended claims. Also, the term "comprising" as used in the specification including the claims is to be construed as meaning "including but not limited to", unless otherwise indicated. As used herein, the terms "a", "an" and "the" are intended to encompass both singular and plural forms, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising", when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. As used herein, the term "and / or" includes any and all combinations of associated items.

[0042] It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting, unless otherwise indicated. As used herein, the terms "fixed" and "connected" and / or "coupled" are used broadly and encompass both direct and indirect connections, and / or coupling, and allow for the potential presence of intermediate device(s) between two elements that are "fixed" and / or "connected" and / or "coupled" together. The term "vertical", "horizontal", "inner", "outer", "left", "right", and similar terms are used for explanation only and are not intended to be limiting.

[0043] For a more complete understanding of the present application, reference is made to the following description taken in conjunction with the accompanying drawings in which:

[0044] Embodiments

[0045] Reference should be made to the drawings for an understanding of the present application, in which: Figure 1 The application provides a Hainan dialect speech recognition optimization system based on knowledge distillation, comprising:

[0046] A data preprocessing module is configured to process Hainan dialect speech data and Hainan dialect text data, and output MFCC speech feature sequences, labeled text labels and Hainan dialect text data.

[0047] A teacher model module comprises an RNN language model, a CNN language model and a Transformer language model, and the dialect text data is input into the RNN language model, the CNN language model and the Transformer language model, respectively, and soft labels and intermediate layer features are obtained through dynamic temperature adjustment.

[0048] The student model training module performs knowledge distillation and parameter tuning on the student model according to the MFCC speech feature sequence, the labeled text label, and the soft label and the intermediate layer feature.

[0049] The output module performs speech recognition on the Hainan dialect by using the student model after knowledge distillation and parameter tuning to obtain a recognition result.

[0050] Specifically, Hainan dialect speech data and Hainan dialect text data corresponding to the Hainan dialect speech data are obtained, and the Hainan dialect speech data and the Hainan dialect text data are input into a data preprocessing module for processing. Specifically, the Hainan dialect speech data is segmented into short-time frames, Fourier transform is performed on each frame of signal to convert from time domain to frequency domain, the power spectrum is calculated, Mel filtering is performed on the power spectrum by a Mel filter bank to obtain a Mel spectrum, logarithmic processing is performed on the Mel spectrum, discrete cosine transform is performed on the logarithmic Mel spectrum, cepstrum coefficients are extracted, and finally an MFCC speech feature sequence is extracted. The Hainan dialect text data is labeled to obtain a text label. Then, the Hainan dialect text data processed by the data preprocessing module is input into a teacher model module, a soft label and an intermediate layer feature are generated after dynamic temperature adjustment, the teacher model module includes an RNN language model, a CNN language model, and a Transformer language model, knowledge distillation and parameter tuning are performed on the student model again, the student model can learn more rich speech structure knowledge by integrating multiple teacher models, a trained student model is finally obtained, and the trained student model is used for speech recognition of the Hainan dialect to obtain an accurate Hainan dialect speech recognition result.

[0051] As an optional implementation, the RNN language model is an LSTM-based recurrent neural network language model, which is used to capture sequence dependency in the Hainan dialect text data.

[0052] The CNN language model is a convolutional neural network-based language model, which is used to capture local features of the Hainan dialect text data.

[0053] The Transformer language model is stacked by multiple Transformer encoders, which is used to extract multi-scale features of the Hainan dialect text data.

[0054] Specifically, the LSTM-based RNN language model can capture the timing characteristics such as rhythm and intonation unique to the Hainan dialect, helping the model better understand the semantic and grammatical structure of the Hainan dialect; the CNN language model based on the convolutional neural network can capture the local features in the Hainan dialect, for example, capture specific syllable combinations and rime combinations in the Hainan dialect, to help the model better identify key features in the Hainan dialect; the Transformer language model can extract multi-scale features in the Hainan dialect, for example, extract features from syllable to sentence level, to help the model better understand the complex structure of the Hainan dialect; by combining the RNN language model, the CNN language model and the Transformer language model, the features of the Hainan dialect can be more comprehensively captured; after weighting and averaging the soft labels generated by the three language models, a more comprehensive supervision signal can be generated to help the student model better learn the knowledge of multiple language models in the teacher model, so that more accurate speech recognition results can be obtained in the subsequent training process.

[0055] Please refer to the accompanying drawings Figure 2 As an optional implementation, the knowledge distillation and parameter tuning of the student model according to the MFCC speech feature sequence, the labeled text label, the soft label and the intermediate layer feature include:

[0056] For the MFCC speech feature sequence, the student model is used for forward propagation to generate a prediction result;

[0057] For the labeled text label, the labeled text label is taken as a true label, and a cross-entropy loss of the prediction result and the true label is calculated to obtain a classification loss;

[0058] For the soft label, the soft labels obtained by the RNN language model, the CNN language model and the Transformer language model are weighted and averaged, and a distillation loss between the prediction result of the student model and the weighted and averaged soft label is calculated;

[0059] For the intermediate layer feature, the real-time intermediate layer feature corresponding to the teacher model in the student model is extracted, the intermediate layer feature output by the teacher model is matched with the real-time intermediate layer feature extracted by the student model, and a cosine similarity loss of the intermediate layer feature of the teacher model and the real-time intermediate layer feature of the student model is calculated to obtain an intermediate layer feature matching loss;

[0060] A total loss function is calculated based on the classification loss, the distillation loss and the intermediate layer feature matching loss, and the trained student model is obtained by training according to the total loss function.

[0061] Specifically, the dynamic temperature adjustment is a method of adjusting the model output probability distribution by introducing a temperature parameter, the original output of the teacher model is the score of each class, which represents the confidence of the input Hainan dialect speech data belonging to each class, the soft label is the probability distribution of the output of the teacher model, the original output of the teacher model is subjected to dynamic temperature adjustment, and a smoother soft label can be generated, and the smooth probability distribution can provide a more fine-grained supervision signal, after weighted average fusion of the soft labels generated by the dynamic temperature adjustment of the RNN language model, the CNN language model and the Transformer language model, multi-dimensional language knowledge can be integrated, and the weighted average soft label is compared with the prediction result output by the student model to calculate the distillation loss, the soft label generated by the RNN language model, the CNN language model and the Transformer language model is used to guide the training of the student model, and the knowledge of multiple language models in the teacher model is fused, so that the generalization ability of the student model can be further improved.

[0062] The traditional fixed temperature cannot balance the global structure learning and the detail discrimination learning, and cannot adapt to the needs of different training stages.

[0063] In the application, the original output of the teacher model is subjected to dynamic temperature adjustment to obtain a soft label, wherein the temperature dynamically changes with the current training round, and the expression is:

[0064] t(e)=t0×exp(-λ×e / E)

[0065] Wherein, E is the total training round, t0 is the initial temperature, the value range is t0∈(0,1], λ is the sensitivity score, the value range is ∈ represents a very small temperature value.

[0066] Exemplarily, if t0=1, λ=2, E=100 rounds, then the temperature changes with the training round as follows:

[0067] Initial temperature: t(0)=1;

[0068] Intermediate temperature (e=50): t(50)=1×exp(-2×50 / 100)=exp(-1)≈0.368;

[0069] Final temperature (e=100): t(100)=1×exp(-2)≈0.135;

[0070] The temperature throughout the whole process is in [0,1], and monotonically decreasing.

[0071] The improved dynamic temperature makes the soft label smoother at the initial stage by using a large t, guides the student model to capture the overall semantic correlation of the language model in the teacher model, and makes the soft label close to the hard label by using a small t at the later stage to strengthen the accurate classification ability (such as the text vocabulary corresponding to a specific voice), so that the training process is more in line with the learning rule, and the recognition accuracy is higher compared with the fixed temperature.

[0072] By calculating the cosine similarity between the intermediate layer features of the teacher model and the intermediate layer features of the student model, the intermediate layer feature matching loss of the intermediate layer features of the teacher model and the intermediate layer features of the student model is calculated, so that the bottom layer semantic representation of multiple language models in the teacher model can be migrated.

[0073] As an optional implementation, the calculation formula of the classification loss is:

[0074]

[0075] wherein y is a real label, t is a temperature, q (t) is a softened probability distribution output by the student model; is a classification loss.

[0076] As an optional implementation, the calculation formula of the distillation loss is:

[0077]

[0078] wherein is is a distillation loss, t is a temperature, p (t) is a softened probability distribution output by the teacher model, M is the number of teacher models, is a temperature-processed soft label of the mth teacher model, w m is a weight;

[0079] The weight calculation formula in the distillation loss is:

[0080]

[0081] Confidence m =1 / CER m

[0082] wherein Confidence m is the confidence of the mth teacher model, calculated by the CER of the teacher model verification set.

[0083] Specifically, the temperature t is used to adjust the "softening" degree of the output; by softening the output through the temperature, the teacher model can deliver more rich implicit information to the student model, instead of relying on the hard label only; the student model can mine the deep knowledge of the teacher model by learning the softened output of the teacher model, and improve its own performance;

[0084] The traditional knowledge distillation method relies on the soft label guidance of a single teacher model (such as an RNN language model), which limits the knowledge source; the present application integrates multiple teacher models (including RNN, CNN and Transformer language models), and designs a confidence-based adaptive weight fusion strategy, that is, the distillation loss is calculated according to the soft label of the multi-language model fusion, so that the student model can learn multi-granularity language features across models, and the student model can learn more rich language structure knowledge.

[0085] As an optional implementation, the calculation formula of the intermediate layer feature matching loss is:

[0086]

[0087] wherein L is the number of intermediate layers, is the feature of the lth layer of the teacher model, is the feature of the lth layer of the student model.

[0088] Specifically, the traditional knowledge distillation method only distills the soft label, and in the present application, the features of the intermediate layers of the teacher model (such as the semantic representation of the language model hidden layer) are simultaneously migrated, so that the student model learns more bottom-level language structure encoding, thereby helping the model to learn more features.

[0089] As an optional implementation, the expression of the total loss function is:

[0090]

[0091] wherein α and β are loss weights, is the total loss, is the classification loss, is the distillation loss, is the intermediate layer feature matching loss.

[0092] Specifically, the total loss function is used for training, so as to integrate external language models in the training process; wherein α and β are loss weights, the range of α and β is preset, the optimal combination is searched in the validation set, and the learning rate is dynamically adjusted using the OneCycle strategy.

[0093] The application significantly improves the accuracy and generalization ability of Hainan dialect speech recognition by introducing the multi-teacher model knowledge distillation and dynamic temperature adjustment strategy. Compared with the traditional single teacher model, the application combines the soft labels of various teacher models such as RNN, CNN and Transformer, enriches the language structure knowledge transmission, and at the same time, the application of the dynamic temperature adjustment strategy makes the training process more in line with the learning rule, which helps the model to obtain good performance in different training stages and optimizes the training process. In addition, the trained student model has small size and low computational complexity, which is convenient for deployment in resource-constrained environments such as embedded devices. The application not only solves the current technical problems, but also provides a new idea for the future development of Hainan dialect speech recognition technology, and is expected to be widely used in tourism, medical care, old-age care and other fields, and make positive contributions to the informatization construction and social development of Hainan area.

[0094] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. The protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A Hainan dialect speech recognition optimization system based on knowledge distillation, characterized in that, The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. wherein, E is the total training epochs, is the initial temperature, with t0∈(0, 1] and λ is the sensitivity score, with , denotes a very small temperature value; The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. wherein, is the distillation loss, t is the temperature, is the probability distribution output by the teacher model after softening, is the number of teacher models, is the soft label of the mth teacher model after temperature processing, is the weight; The application relates to a speech recognition method and device based on a teacher-student model. wherein, is the confidence of the mth teacher model; The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. wherein L is the number of intermediate layers, is a feature of the teacher model at layer, is a feature of the student model at layer; The application relates to a speech recognition method and device based on a teacher-student model.

2. The knowledge distillation based Hainanese dialect speech recognition optimization system according to claim 1, wherein, The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. 3.The knowledge distillation based Hainanese dialect speech recognition optimization system according to claim 1, wherein, The application relates to a speech recognition method and device based on a teacher-student model. where, is the true label, is the temperature, is the student model output softened probability distribution; is the classification loss.

4. The knowledge distillation based Hainanese dialect speech recognition optimization system according to claim 1, wherein, The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher-student model. The application relates to a speech recognition method and device based on a teacher wherein, , is a loss weight, is a total loss, is a classification loss, is a distillation loss, is an intermediate layer feature matching loss.

Images