End-to-end speech recognition method and system and storage medium

Abstract

The invention provides an end-to-end speech recognition method and system. The method comprises the following steps: training an initial speech feature extraction model by using a source corpus based on a VGGNet model; removing a full connection layer in the initial speech feature extraction model and freezing a preset number of convolutional layer parameters, and training the removed and frozen initial speech feature extraction model by using a target corpus to obtain a frequency domain feature extraction network; constructing an end-to-end speech recognition framework, wherein the framework comprises an encoder and a decoder; and training the end-to-end speech recognition framework by using a target corpus, and performing end-to-end speech recognition based on the trained end-to-end speech recognition framework. The method can effectively solve the problem of model overfitting under the condition of limited data, improves the accuracy of speech recognition, and has good noise robustness.

Description

technical field [0001] The invention belongs to the field of speech recognition, and in particular relates to an end-to-end speech recognition method, system and storage medium. Background technique [0002] In the past few decades, speech recognition technology has made great progress driven by artificial intelligence technology. End-to-end speech recognition based on deep learning can directly complete speech-to-text transcription through a single model, and its performance exceeds Compared with traditional speech recognition methods, it reaches the state-of-the-art in some tasks. Therefore, end-to-end speech recognition has received extensive attention from academia and industry. [0003] In traditional speech recognition, the most common algorithm is Hidden Markov Model (GMM-HMM) based on Gaussian Mixture Model. With the rise of deep neural network (DNN), the DNN-HMM framework uses the DNN model to replace the original GMM model to model each state without making assum...

AI Technical Summary

Problems solved by technology

However, in the case of limited data (the amount of training data is less than 100 hours), the model is fully trained due to insufficient data, the model is overfitting, the generalization is poor, and the robustness to noise is poor, and its recognition effect is not as good as that of Traditional Speech Recognition Technology

Although the time-frequency information of speech signals can be better mined by introducing more and more complex network computing layers, more parameters are introduced at the same time, and the knowledge learned by the model is limited when the data is limited. cause the model to fit the data bias

Images