Speech Recognition System of Specific Vocabulary

End-to-end speech recognition system can realize the conversion from input audio to character text without complicated data processing, and it is widely used in speech recognition field. In this paper, the speech recognition system of DeepSpeech2 is built by combining CNN network which is good at capturing sequence spatial features and RNN network which is good at capturing time sequence information. The system is trained based on the THCHS-30 Dataset, and achieves more accurate speech prediction by simplifying the process and enhancing the use of context information. Experiments based on this dataset show that the word error rate of this model is only 0.133, and the effect is very ideal.