Simultaneous model re-estimation from contaminated data by composed hidden Markov modeling

Abstract
The problem of estimating speech models from noisy data is considered as a generalization of the Baum-Welch reestimation algorithm. The general approach to this problem is pursued by considering the interaction of speech data frames with noise data frames produced by independent speech and noise sources. It is shown that the generalization of the Baum-Welch reestimation formulae can be used to estimate the speech and noise models from contaminated data. The performance of the estimated models is evaluated for recognition in quiet and noisy environments. The background noises investigated are stationary pink noise and impulsive machine gun bursts.

This publication has 3 references indexed in Scilit: