Third International Conference on Spoken Language Processing (ICSLP 94)
This paper discusses a highly reliable phoneme recognition method in various styles of utterance based on mutual information criterion. Mutual information is a good measure to build an effective phoneme dictionary in the process of optimal selection of acoustic features and integration of clusters. Using VQ code sequences organized by the hierarchical clustering method, phonemic likelihoods for ea,ch frame can be calculated. Phoneme recognition is performed with applying phonemic duration and bigram constraints of phonemes. Also, we cover an iterative training mechanism of the phoneme dictionary. The correct rate for phoneme is improved to 90.5% (8.4% insertion, 7.0% deletion) in the speaker independent recognition experiment for the continuous utterance.
Bibliographic reference. Okawa, Shigeki / Kobayashi, Tetsunori / Shirai, Katsuhiko (1994): "Phoneme recognition in various styles of utterance based on mutual information criterion", In ICSLP-1994, 1911-1914.