Recognition of Dialogue Speech
Shozo MAKINO, Akinori ITO,
Takashi Otsuki, Yukihiro Osaka, Shinkou AWATSU, Yoshiyuki OKIMOTO
Department of System Information Science,
Graduate School of Information Sciences, Tohoku University
Katahira 2-1-1 SKK Build., Aoba-ku, Sendai, 980 Japan
e-mail: makino@rcais.tohoku.ac.jp
We have carried out the research for recognition of dialogue speech. The research is separated into two parts. One is to construct a robust phoneme recognition method with high accuracy. Another is to develop a language processing system for spoken dialogue. For the phoneme recognition, we proposed two methods: (1) a unified system connecting the DP-based time warping with the discriminant training method. (2) a speech recognition system with adaptability function to speech rate of input speech using preliminary phoneme recognition results. For the language processing, we have been building up a system using a dependency grammar based on Bunsetsu-units. Now we have modified the Bunsetsu automaton so as to accept various Bunsetsu units appeared in spoken dialogue. Furthermore we developed a high speed method for the Bunsetsu-unit detection, and a performance prediction method based on number of sentences with a short distance.
Keywords: Discriminant training, Adaptation to input speech, Grammar for spoken dialogue, High speed Bunstsu detection, Performance prediction