Acoustic models adaptation in large vocabulary continuous Mandarin speech recognition for non-native speakers

In this paper, a number of acoustic modeling techniques are implemented to compare their performance on non-native speech recognition with the linguistic minorities accents Mandarin speech corpus, which collected, by our lab. Firstly, we train native baseline HMM models using the project 863 standard Mandarin corpus. Secondly, with initial parameter values of the native baseline models the non-native accent-dependent HMM models are trained using the speech spoken by the speakers from Naxi and Lisu in Yunnan, China. Furthermore we use the MLLR to do the speaker adaptation experiments. It is shown that when accent-dependent HMM and MLLR are used, the error rates reduce evidently.