Speech recognition in nonstationary noise based on parallel HMMs and spectral subtraction

This paper proposes a method of speech recognition in a nonstationary noisy environment, combining the parallel HMMs and the spectral subtraction. In the proposed method, a set of hypothesis is generated with respect to the combination of the speech and the noise that can produce the observed data by a series of subtraction processes. Using HMMs prepared separately for the speech and the noise, the probabilities of occurrence are calculated. The 100-word recognition in the noisy environment in an ordinary car running in an urban area, is defined as the task in the experiment. Comparative experiments, are made for the proposed method, the ordinary spectral subtraction method and other parallel HMM methods. Then, the effectiveness of the proposed method is verified.