Multi-path Syllable Models Based on Phonetic Knowledge

Recent research suggests that syllable-length acoustic models might be more appropriate for pronunciation variation modelling than the context-dependent phones that conventional automatic speech recognisers use. In this paper, we compare the recognition performance of two types of recognisers: a conventional recogniser that only uses triphones, and an experimental recogniser that employs a mix of context-independent syllable models for a set of frequent syllables and sequences of triphones for the less frequent ones. The syllable models of the mixed-model recogniser are designed to consist of multiple HMM paths that are expected to capture major pronunciation variants. These paths are initialised using phonetic knowledge and re-estimated using a data-driven solution. When applied to 94 frequent syllables in a 37-hour corpus of Dutch read speech, the multi-path mixed-model recogniser outperforms a much more complex triphone recogniser.