On the effectiveness of speech enhancement to a proposed speech recognition process that applied to noisy isolated-word recognition

A speech recognition process is proposed and applied to noisy speech recognition. The end points of isolated utterance can be determined accurately enough in the training and recognition process. A function for the distance measure between two speech blocks is derived. The feature vector is then defined. The speech model for each word is represented as a left-to-right finite-state machine. Viterbi algorithm and EM algorithm is applied to obtain the optimal model for each isolated word. When an utterance is to be recognized, the Viterbi algorithm is applied to search for the model which gives the smallest accumulated distance along the optimum path. The performance of the proposed recognition process is evaluated by comparing to the method using DTW. The comparison is in three aspects: namely, the recognition accuracy, the noise-resistance capability, and the effectiveness of speech enhancement pre-processing. Experiments showed that the proposed speech recognition process is superior to DTW in these three aspects.