A study on Yunnan dialectal Chinese speech recognition

In this paper, a Yunnan dialectal accents Putonghua speech corpus which collected by our lab is used in Yunnan dialectal Chinese speech recognition study. Several acoustic adaptation techniques are preformed to improve the recognition rate for accented speech and their performances are compared. Firstly, the baseline system is trained using the project 863psila standard Putonghua corpus, the recognition rate of Yunnan dialectal accents speech is obtained. It is lower than the recognition rate of standard Putonghua speech. Secondly, based on the Yunnan dialectal accents Putonghua speech corpus, Yunnan dialectal accents acoustic models are built, their performances are tested. Finally, speaker adaptation experiments are implemented. MLLR and MAP adaptation methods are used. The results show the proposed methods can evidently improve the recognition performances for Yunnan dialectal accents.