A Streaming End-to-End Speech Recognition Approach Based on WeNet for Tibetan Amdo Dialect

Speech recognition is a technique to transcribe acoustic features into text sequences. However, traditional speech recognition model cannot get an effective performance, when dealed with Tibetan Amdo dialect dataset which requires a large amount of linguistic knowledge. In order to solve this issue, we propose an end-to-end speech streaming recognition model which can not only realize the transcription of the Tibetan Amdo dialect but also solve the Tibetan Amdo dialect streaming recognition problem. In the model, we choose Tibetan syllables as the modeling unit and MFCC as the acoustic features. Furthermore, extensive experimental results show good results on our self-built thousand-hour level dataset. Finally, the Character Error Rate(CER) of speech streaming recognition on ourself dataset is 10.73%, which is a relative improvement is 17.65% compared to the baseline model. The CER of speech transcription is 10.23%, which is a relative improvement is 16.01% compared to the baseline model.