A TONE RECOGNITION FRAMEWORK FOR CONTINUOUS MANDARIN SPEECH

In this paper, we present a tone recognition framework for continuous Mandarin speech. To model the variations of F0 pattern caused by co-articulation and phonetic effects, a set of discriminating features are extracted: 1) outlined features from the F0 contours of target syllable and neighboring syllables are combined; 2) contextual tone information is utilized within an iterative process; 3) phonetic information from target and neighboring syllables is incorporated. These features are put into a decision tree for tone classification, which follows an HMM-based toneless decoder. The results in 5-tone recognition experiments show more than 40% relative error rate reduction against the baseline local outlined features. Moreover, the proposed method obviously outperforms HMM-based tone model in speaker-independent evaluation.