In this thesis work, the methods for robust speech recognition in noisy environment are developed by adapting hidden Markov model(HMM) parameters. The adaptation can be accomplished by modifying the means and variances of the HMM's continuous observation parameters according to the correction factors that map the Gaussian distribution of clean speech model to the distribution of noisy speech model.
We obtain the formulation of correction factors using maximum a posteriori (MAP) estimation that is useful when only sparse noisy training data are available. We develop vector field interpolation(VFI) method which significantly reduces the training data size and stabilizes the adaptation procedure compared to the case of MAP method only. Because a few Gaussian mixtures contribute the observation probabilities mainly, the correction factors corresponding to these mixtures are used to adjust all the other factors to the same direction of these main factors in VFI.
In the isolated-word recognition experiments based on semicontinuous HMM and MAP/VFI adaptation method, we improve recognition performance considerably in noisy environment.
