Prosody-dependent Acoustic Modeling for Mandarin Speech Recognition

Abstract

A study on introducing prosodic information to acoustic modeling (AM) for speech recognition is reported in this paper. It extends the conventional context-dependent (CD) triphone HMM modeling approach to further consider the dependency of phone model on the break type of nearby inter-syllable boundary. Four break types are considered, including major break, minor break, normal non-break, and tightly-coupled non-break. In the training phase, break labeling is automatically accomplished by a Prosody Labeling and Modeling algorithm proposed previously. Then, prosody-and phonetic-dependent phone models are constructed by a standard decision tree-based context clustering of HMMs. The effectiveness of the new AM was examined on a Mandarin syllable recognition task. Experimental results showed that the new approach outperformed the conventional CD-AM on achieving better syllable recognition rate as well as on obtaining a more efficient syllable lattice with better compromise on complexity verse syllable coverage rate.