Bayesian learning of Gaussian mixture densities for hidden Markov models

摘要: An investigation into the use of Bayesian learning parameters a multivariate Gaussian mixture density has been carried out. In continuous hidden Markov model (CDHMM) framework, serves as unified approach for parameter smoothing, speaker adaptation, clustering, and corrective training. The goal this study is to enhance robustness in CDHMM-based speech recognition system so improve performance. Our incorporate prior knowledge CDHMM training process form densities HMM parameters. theoretical basis procedure presented preliminary results applying clustering are given.Performance improvements were observed on tests using DARPA RM task. For under supervised mode with 2 minutes speaker-specific data, 31% reduction word error rate was obtained compared speaker-independent results. Using Baysesian smoothing sex-dependent modeling, 21% FEB91 test.