ASR2000 - Automatic Speech Recognition: Challenges for the new Millenium September 18-20, 2000 Paris, France

Bayesian Adaptation Revisited

Patrick Kenny, Gilles Boulianne, and Pierre Dumouchel

Centre de recherche informatique de Montréal, Montréal, Québec, Canada

We report the results of some preliminary experiments with a new method of acoustic-phonetic modeling for large vocabulary applications that can be viewed as a far-reaching extension of Bayesian speaker adaptation. This method adapts all of the Gaussian mean vectors in a speaker-independent HMM for a given speaker (and not just the mean vectors present in the speaker’s adaptation data as in classical Bayesian adaptation). It is based on an explicit model of the correlations between all of the speakers in the training set, the idea being that if there is not enough data to estimate a Gaussian mean vector for a given speaker then data from other speakers can be used provided that we know how the speakers are correlated with each other. Our new approach has resulted in 10-15% reductions in error rate on a French language dictation task.

Full Paper (PDF)   Full Paper (Zipped Postscript)

Bibliographic reference.  Kenny, Patrick / Boulianne, Gilles / Dumouchel, Pierre (2000): "Bayesian adaptation revisited", In ASR-2000, 112-119.