This paper presents an empirical study of word error minimization approaches for Mandarin large vocabulary continuous speech recognition (LVCSR). First, the minimum phone error (MPE) criterion, which is one of the most popular discriminative training criteria, is extensively investigated for both acoustic model training and adaptation in a Mandarin LVCSR system. Second, the word error minimization (WEM) criterion, used to rescore N-best word strings, is appropriately modified for a Mandarin LVCSR system. Finally, a series of speech recognition experiments is conducted on the MATBN Mandarin Chinese broadcast
news corpus. The experiment results demonstrate that the MPE training approach reduces the character error rate (CER) by 12% for a system initially trained with the maximum likelihood (ML) approach. Meanwhile, for unsupervised acoustic model adaptation, MPE-based linear regression (MPELR) adaptation outperforms conventional maximum likelihood linear regression (MLLR) in terms of CER reduction. When the WEM decoding approach is used for N-best rescoring, a slight performance gain over the conventional maximum a posteriori (MAP) decoding method is also observed.l