Interspeech'2005 - Eurospeech

Lisbon, Portugal
September 4-8, 2005

Environmental Compensation Using ASR Model Adaptation by a Bayesian Parametric Representation Method

Xuechuan Wang, Douglas O'Shaughnessy

Université du Québec, Canada

The mismatch between system training and operating conditions can seriously deteriorate the performance of ASR systems. The maximum a posteriori (MAP) estimation is used for the adaptation of HMM-based multivariate Gaussian mixture models (GMMs). In this paper, we propose an environment independent ASR model parameter adaptation approach based on Bayesian parametric representation (BPR). Compared to the MAP method, the BPR adaptation method has better performance with limited adaptation data. The performances of the two methods are investigated in the experiments designed on the AURORA 2 noisy speech database.

Full Paper

Bibliographic reference.  Wang, Xuechuan / O'Shaughnessy, Douglas (2005): "Environmental compensation using ASR model adaptation by a Bayesian parametric representation method", In INTERSPEECH-2005, 1801-1804.