Syllable recognition for continuous Japanese speech recognition

This paper presents a syllable recognition method for continuous Japanese speech which are naturally and fluently spoken. In the method, syllable matching was made using both vowel templates (stationary patterns) and syllable templates (transitional patterns) at the hypothesized VCV sections based on DP matching. The templates were automatically created from a set of continuous utterances. The syllable templates were intended to contain only the consonant and its neighbouring portion as much as possible, in order to cope with variations, deletion or shortening of vowels due to the coarticulation effects of phonemes. The recognition result was obtained as a form of syllable lattice, which was fed into a linguistic processor[5]. As the results of experiments using four talkers´ 200 phrases, 1.3 syllable sections were hypothesized for a syllable on the average and a syllable recognition rate of 79.3% (96.6% in top ten answers) was obtained.