Word-Level Tone Modeling for Mandarin Speech Recognition

Standard HMM-based Mandarin speech recognition systems do not exploit the suprasegmental nature of tones, but explicit tone models can be incorporated with lattice rescoring. This work extends previous approaches to explicit tone modeling from the syllable level to the word level, incorporating a hierarchical backoff. Word-dependent tone models are trained to explicitly model the tone coarticulation within the word. For less frequent words, tonal-syllable-dependent tone models or plain tone models are used as backoff. More generally, context-dependent tone models can be used as backoff. The word-dependent tone modeling framework can be viewed as a generalization of the traditional context-independent and context-dependent tone modeling. Under this framework, different types of tone modeling strategies are compared experimentally on a Mandarin broadcast news speech recognition task, showing significant gains from the word-level tone modeling approach.