Robust Speech Recognition in Reverberant Environment by Optimizing Multi-band Spectral Subtraction

Reverberant environment poses a problem in speech recognition application where performance degrades drastically depending on the extent of reverberation. Thus, it is important to employ front-end speech processing, such as dereverberation to minimize its effect. Most dereverberation techniques used to address this problem enhance the reverberant waveform prior to speech recognition. Although the speech quality is improved, this approach treats the front-end speech enhancement and the recognizer independently. In this paper, we present an approach that treats both dereverberation and speech recognition interdependently. In our proposed approach, the dereverberation parameters are optimized to improve the likelihood of the acoustic model. The system is capable of adaptively fine-tuning these parameters jointly with acoustic model training. Additional optimization is also implemented during decoding of the test utterances. Experimental results show that the proposed method significantly improves the recognition performance over the conventional approach with a relative improvement of 5%.