Optimal SNR Model Selection in Multiple-Model Based Speech Recognition System

In the multiple-model based speech recognition system, multiple HMM models corresponding to different types of noise signals and SNR values are trained and the one model which is most close to the input speech is selected for recognition. In the previous research on the multiplemodel based speech recognition, it has been thought that the best performance can be obtained by selecting the HMM model which is most similar in SNR values to the input speech. But, from our experimental results, it has been found that better performance can be obtained when there is some mismatch between the SNR values of input speech and the selected HMM model. In this paper, we experimentally determined the optimal HMM models corresponding to the SNR values of the input speech in the multiple-model based speech recognizer. From the recognition experiments on Aurora 2 database, we could see far better recognition results compared with the conventional multiple-model based speech recognizer by using the experimentally determined optimal HMM models.