FRAME-BY-FRAME HMM ADAPTATION FOR REVERBERANT SPEECH RECOGNITION

This paper describes a technique for robust speech recognition under a reverberant environment. Acoustic signals are distorted by their reflection inside a room, and this reverberation degrades the performance of automatic speech recognition system. When the effect of reverberation spans for several frames, dealing such the distortion due to reverberation simply as convolutive distortion is not sufficient. In such cases, it becomes necessary to consider the reverberation effect of other frames over current frame. In this paper, frame-synchronous estimation of reverberation component and HMM combination are proposed to deal with such distortion. The proposed technique was evaluated with an isolated word reverberant speech recognition task, and the result showed improvement in recognition performance. Moreover, use of Jacobian adaptation for a frame-synchronous model adaptation is investigated.