Interspeech'2005 - Eurospeech

Lisbon, Portugal
September 4-8, 2005

Fast Unsupervised Speaker Adaptation Through a Discriminative Eigen-MLLR Algorithm

Bart Bakker, Carsten Meyer, Xavier Aubert

Philips Research Laboratories, Germany

We present a new method for unsupervised, fast speaker adaptation that combines the Eigen-MLLR transform approach with discriminative MLLR. We thereby aim to profit both from the performance improvements that are generally provided by a discriminative approach, and from the reliability that Eigen-MLLR has demonstrated in fast adaptation scenarios. We present first evaluation results on the Spoke 4 subset of the 1994 Wall Street Journal (WSJ) database. Our results show that, in fast enrollment scenarios, discriminative Eigen-MLLR allows for clear improvements both over non-discriminative Eigen-MLLR and over discriminative MLLR. We further introduce a method to estimate the weight parameters of Eigen-MLLR discriminatively, and show that this allows for further improvements on the considered data sets.

Full Paper

Bibliographic reference.  Bakker, Bart / Meyer, Carsten / Aubert, Xavier (2005): "Fast unsupervised speaker adaptation through a discriminative eigen-MLLR algorithm", In INTERSPEECH-2005, 257-260.