Integration of tonal knowledge into phonetic HMMs for recognition of speech in tone languages

摘要

A method to integrate tonal knowledge into phonetic hidden Markov models (HMMs) for use in recognition of speech in tone languages is proposed. Tone segment is introduced, and each phonetic HMM models both the phoneme and the associated tone segment. The observation probability density function (pdf) of each HMM state is the product of a phonetic feature pdf and a tonal feature pdf Trainability of the HMMs is maintained by sharing of tonal feature pdrs among different HMMs that model the same tone segment and by sharing of phonetic feature pdrs among HMMs of same phoneme. A simple method that effectively avoids performance degradation due to harmonic pitch errors is also introduced. Results from speaker-independent connected syllable recognition experiments show that the proposed method effectively integrates tonal knowledge into phonetic HMMs.