Conversational Speech Recognition In Non-Stationary Reverberated Environment

This paper presents a conversational speech recognition system able to operate in non-stationary reverberated environments. The system is composed of a dereverberation front-end exploiting multiple distant microphones, and a speech recognition engine. The dereverberation front-end identifies a room impulse response by means of a blind channel identification stage based on the Unconstrained Normalized Multi-Channel Frequency Domain Least Mean Square algorithm. The dereverberation stage is based on the adaptive inverse filter theory and uses the identified responses to obtain a set of inverse filters which are then exploited to estimate the clean speech. The speech recognizer is based on tied-state cross-word triphone models and decodes features computed from the dereverberated speech signal. Experiments conducted on the Buckeye corpus of conversational speech report a relative word accuracy improvement of 17.48% in the stationary case and of 11.16% in the non-stationary one.