Automatic syllabification and phoneme class labelling with a phonologically based hidden Markov model and adaptive acoustical features

In this paper, a method is proposed for segmenting syllables and syllable classes in continuous speech. The implementation involves a combination of phonological knowledge, speaker adaptable features, vector quantization and a hidden Markov modelling technique. Phoneme classes are represented as hidden Markov model states while syllable related phonological knowledge defines the state connections. The observation sequence is distances obtained from the decision planes of two single layer neural networks, which are respectively used to make voiced/unvoiced and (syllabic) nucleus/non-nucleus classifications. Tests were performed on a (bandlimited FM broadcast speech quality) database, containing speech from 10 male speakers. Our results indicate that 97·1% of all syllables were detected and 96·5% phoneme classes were labelled correctly.