In this paper we propose a new approach of speech recognition based on an idea of Statistical Phoneme Center. The Statistical Phoneme Center has several properties that are feasible to realize a high-reliable phoneme extraction. First, in every phoneme we assume that there is the fictitious center which can be estimated statistically. The center is determined by an iterative procedure to maximize the local likelihood using a large amount of speech data. Next, HMMs are trained for every segment between the centers and are concatenated to make a word HMM. The word recognition is realized by an optimal algorithm considering the center likelihood. As the experimental result, 97.7% recognition accuracy is obtained for 216 word vocabularies in the speaker independent condition. The result demonstrates the effectiveness of the proposed method.
Bibliographic reference. Okawa, Shigeki / Shirai, Katsuhiko (1995): "Estimation of statistical phoneme center and its application to accurate phoneme modelling", In EUROSPEECH-1995, 791-794.