INTEGRATION OF TONE RELATED FEATURES FOR MANDARIN SPEECH RECOGNITION BY A ONE-PASS SEARCH ALGORITHM

How to model Chinese tones and integrate them into an HMM-based recognizer for Chinese recognition has long been an area of interest to researchers. In this paper, we pro-pose the use of a polynomial trajectory model to represent pitch shape. We further propose an efficient one-pass search approach that integrates the tone likelihood into the Viterbi search procedure. We report a number of experimental re-sults on tone classification and tonal syllable recognition in the 863 corpus. While the improvement in the tonal syllable accuracy is small, it nevertheless shows the feasibility of the proposed approaches. 1