Abstract
Noise reduction has been extensively studied to prevent the accuracy of automatic speech recognition (ASR) being severely degraded due to various noise sources observed in practical use. The effect of noise reduction is usually dependent on the parameters used in the noise-reduction method, which need to be tuned for different types of noise to achieve the best accuracy. We propose a method for automatically switching noise-reduction parameters in such a way that ASR accuracy is maximized. The proposed method assigns the most suitable parameter set to each speech sentence by measuring and grouping the characteristics of noise observed in each sentence. Experiments using speech sentences contaminated by various types of noise showed that the proposed method can reduce the word error rate of ASR.