Speaker adaptation in noisy environments based on parameter estimation using uncertain data

This paper describes new method for the speaker adaptation of HMM parameters in environments with background noise. This method is based on Bayesian estimation, and calculates the a posteriori distribution of cleanspeech HMM parameters from their a priori distribution by using noisy speech observations. The advantage of the method is that the distribution of the noise can be taken into account in adapting clean-speech HMMs to a target speaker’s speech without noise. The results of the experiments using noninformative prior show that the recognition performance in a noise-free environment was improved by this method even when the SNR of the noisy speech data used for the adaptation was −6 dB.