Application of Deep Learning in Chinese Speech Recognition System

In order to study the application of deep learning in Chinese speech recognition systems, the author introduces the steps of data processing and feature extraction, the network structure, and creation process of acoustic model based on LSTM-CTC and BiLSTM-CTC and describes the optimization strategy used in training acoustic model. This paper describes the extraction methods of Fbank features and MFCC features, then proposes to use LeakyReLU function as the activation function, and analyzes the feasibility by comparing with ReLU function. This paper introduced the hardware configuration of the experimental platform and the deep learning framework TensorFlow and built a speech recognition system based on this framework. Different experiments have been carried out on the acoustic model created, the experiments have verified that the LSTM structure has a significant improvement compared with the RNN structure, among them, BiLSTM-CTC has better performance, and using Fbank features to train the model can improve the recognition accuracy of the system. At the same time, the LeakyReLU function and ReLU function can further improve the performance of the system. The research results found that the word error rate of the Chinese speech recognition system based on the BiLSTM-CTC model is 11.87%, which is significantly lower than other models. In addition, replacing the ReLU function with the LeakyReLU function as the activation function of BiLSTM-CTC can improve the accuracy of the system, with a word error rate of 11.08%.