Research on Improving Speech Recognition System for Oral English Evaluation using Deep Learning Model

This paper focuses on speech recognition algorithm based on bidirectional long short-term memory network combined with Connected Temporal classification (BiLSTM-CTC), aiming to improve the performance of spoken English evaluation system. The algorithm can improve the accuracy and robustness of speech recognition effectively by analyzing the English speech signals after pre-emphasis processing. Firstly, the original speech signal is preweighted to enhance the high frequency part and reduce the frequency resonance effect. Then Mayer frequency cepstrum coefficient (Fbank) is used as feature extraction method to capture the spectral characteristics of speech signals. The experimental results show that the amplitude of energy change of English speech signal after pre-accentuated is stable between-35dB and OdB, and the spectrum distribution is more balanced. This model has a significant improvement in recognition speed and accuracy. Especially in the complex background noise environment, the anti-interference ability of the new model is particularly outstanding, and the false recognition rate is effectively reduced. The speech recognition algorithm based on BiLSTM-CTC proposed in this paper not only optimizes the speech signal processing flow, but also greatly improves the overall performance of the recognition system.