Recognition of continuous utterances of Japanese vowel sequences based on structural representation of speech

Non-linguistic features such as vocal tract shapes and acoustic devices are inevitably involved in speech. Recently, a new representation of speech without any dimensions indicating the non-linguistic features was proposed. It discards the absolute properties of speech events and captures only the interrelations among them. In this paper, recognition experiments of continuous utterances of Japanese vowel sequences using stractural representation of speech are carried out. Distribution sequences are obtained from continuous utterances by training HMMs, and ML-based, MAP-based and Bayes-based methods are investigated for the estimation of the HMM parameters. Furthermore, to solve problems in continuous utterances recognition, a DP-based structural matching method is proposed and examined in recognition experiments.