A speech processing based on syllable identification by using phonological patterns

In this paper we describe a speech recognition method including syllable identification and syllable hypothesization in both which phonological variations have been taken into account. The syllable identification is carried out by matching a sequence of phonemically characterized segments with syllable patterns written in a descriptive form. The syllable patterns involve phonological variants (e.g., vowel reduction or vowel deletion) within a syllable unit. Syllable hypotheses are generated from original entries, including syllable substitution or syllable deletion caused by phonological influence. The preliminary recognition experiments are carried out by applying the recognition method to isolated word recognition using 300-word utterances of 10 male and 10 female adults. The results show recognition rates of 91- 94% for 300 words after training using small subset (36 words) selected from the 300 words. The approach presented here has the potential of being available in representing and recognizing continuous speech.