The improvements on automatic mandarin pronunciation evaluation
Yi He, Yunpeng Xiao, Zhao Xinru, Xin Qi, Xiaoying Xu, Weiping Ye · 2012
Automatic pronunciation evaluation (APE) often works in noise environment. With MFCC based automatic speech recognizer (ASR) embedded, APE's noise robustness is not usually satisfied. Since it improves anti-noise performances of ASR, power normalized cepstrum coefficient (PNCC) [1] is introduced i n t o APE in this paper. Experiment of noise environment shows improvement in both APE and recognition performances. In order to discriminate mispronunciation APE needs standard “golden” acoustic models. The speaker adaptation models of APE need to be trained by standard pronounced utterances of target speaker. But standard pronunciation is not usually expected in APE. In this paper a new speaker adaptation strategy is suggested. A recognizer of speaker independent (SI) is used to selected correct pronounced utterances for adaptation training. The adapted system achieved better score correlation with subjective scores.