Speech Recognition Under the Unstationary Noise Based on the Noise Markov Model and Spectral-Subtraction

In this paper, mechanisms to recognize speech in time varying noise are proposed. In these methods, the noise source and the speech source are modeled independently. The probabilities that these two models emit observed data sequence are calculated. We tested two methods. The first one uses HMMs to model both of noise and speech. The second one uses a normal Morkov model for noise representation. Adopting normal Markov model, the noise model itself become rather complex, however, the dynamical features such as delta cepstrum can be easily and precisely considered. Using these methods, 100 word recognition tests in car noise environment are performed. As the results, the performances are improved by 23 % and 26 % using the first and the second method, respectively, as compared with normal spectral subtraction.