Speaker Adaptation Using Lattice-based MLLR

This paper presents lattice-based maximum likelihood linear regression (MLLR) for unsupervised adaptation. Lattice MLLR accumulates the statistics used in the MLLR transform estimation procedure using a forward-backward pass through a word-lattice of alternative hypotheses rather than assuming that the 1-best transcription is accurate as in standard unsupervised MLLR. This results in the ability to robustly estimate a larger number of transforms from the same amount of adaptation data. Lattice-based MLLR can therefore yield lower word error rates than standard unsupervised MLLR and it is compared experimentally to a version of MLLR using confidence scores. Recognition experiments show that lattice-based MLLR can reduce word error rates on Switchboard Minitrain task by 1.4% absolute and on NIST Hub5 1998 evaluation set by 1.0%.