Optimizing feature extraction for English word recognition

In this paper, we investigate the possibility to improve the recognition accuracy of English words by optimizing feature extraction. The optimization method is based on the simplex algorithm. Although the mel-cepstrum obtained by the critical band filters reflects the human auditory perception, it may not be the optimal features for speech recognition. In this paper, we show that the mel-cepstrum can be optimized in terms of recognition accuracy by adjusting the center frequencies and bandwidths of the critical filters. Experiments with English words showed that the optimized filter bank provide a noticeable performance improvement.