A robust front-end for telephone speech recognition

In this study, we propose an effective front-end technique to improve the performance of telephone speech recognition. Many works have been concentrated on compensating the noise and the channel distortions contained in telephone speech at the front-end stage of speech recognition. Based on RASTA processing which is well known for its channel robust feature parameters, we tried to further improve this method using the channel estimation power of cepstral mean subtraction and maximum likelihood method. As a hybrid method of channel estimation and RASTA processing, the proposed method was proved to be effective by experiments performed on real telephone speech data.