Sampling Methods for Instantaneous Speaker Adaptation

Abstract Linear mean-based adaptation schemes, in which an ane linear transformation is ap- plied to the mean of all Gaussian components in an LVCSR system, have been very successful as an approach to speaker adaptation. In particular, Maximum Likelihood Linear Regression (MLLR) oers,a simple EM approach to,nding the maximum like- lihood (ML) transform. However, robustness issues arise when using the ML transform to adapt to a new speaker on very little data, due to the fact that the transform poste-