IMPROVE SPEECH RECOGNITION PERFORMANCE IN REVERBERANT ENVIRONMENT BASED ON ESTIMATION OF ENERGY FEATURE

Abstract

The purpose of this paper is to improve speech recognition performance in reverberant environment with distant talking,which based on directly speech processing or selected Log-energy feature. Speech recognition performance can be improved by changing the value of Log-energy feature or directly speech energy. Experiments used CENSREC-4 corpus to evaluate distant-talking speech under various reverberation environments by single microphone. And likelihood score string is proposed to optimal Energy feature of speech for each utterance to improve recognition performance. The results have showed improvements of speech recognition performance in reverberant environments with our proposals.