Improved acoustic modeling with Bayesian learning

The authors study the use of Bayesian learning for the estimation of the parameters of a multivariate mixture Gaussian density. For speech recognition algorithms based on the continuous density hidden Markov model (CDHMM) framework, Bayesian learning serves as a unified approach for the following four applications: parameter smoothing, speaker adaptation, speaker group modeling, and corrective training. In the approach, the authors use Bayesian learning techniques to incorporate prior knowledge into the CDHMM training process in the form of prior densities of the HMM parameters. The theoretical basis for this procedure is presented. All four applications have been evaluated. Experimental results of the TI connected digit task and the Naval Resource Management task are provided to show the effectiveness of Bayesian adaptation of CDHMM.<<ETX>>