The Study of Prosodic Modeling for Mandarin Speech

Abstract

Prosodic modeling plays an important role in the integration of speech recognition and natural language understanding. The function of prosodic modeling for the integration of speech recognition and natural language understanding is to explore the prosodic phrasing of the testing utterance for providing useful information to help linguistic decoding in the next stage. The main concern of the prosodic phrasing issue is to build a model describing the relationship between input prosodic features extracted from the testing utterance and output linguistic features of the associated text. In this paper, three prosodic modeling approaches based on the VQ, SOFM, and RNN techniques have been discussed in detail. Experimental results have shown that they all functioned well on detecting prosodic states from acoustic cues. By evaluating on perplexity reduction, we found that the RNN-based approach outperformed the other two approaches.