Improved speaker adaptation using speaker dependent feature projections

We extend the formulation of constrained maximum likelihood linear regression (CMLLR) adaptation to take into account full covariance matrices in the adapted model, and we use it in conjunction with heteroscedastic linear discriminant analysis (HLDA) in order to estimate speaker dependent feature projections on both training and test data. Results on the broadcast news corpus show that the proposed HLDA adaptation technique is very effective, even when combined with traditional CMLLR and MLLR adaptation, providing up to 8% relative improvement in recognition accuracy.