Very fast adaptation for large vocabulary continuous speech recognition using eigenvoices

The principle of the eigenvoice method | using a priori knowledge on the speaker variability as collected during the training for a very fast adaptation | is applied to continuous speech recognition with large vocabulary. The handling of mixture density HMMmodels is discussed. For the case of gender independent models, a decrease of the word error rate of up to 15% is observed for unsupervised adaptation and even the rst recognized phonemes lead to considerable improvements. The rst two eigenvectors of adaptation can be characterized as classifying the gender and the recording environment. Comparisons of the method with MLLR are done as far as the latter is appli-