Iterative unsupervised adaptation using maximum likelihood linear regression

Maximum likelihood linear regression (MLLR) is a parameter transformation technique for both speaker and environment adaptation. In this paper, the iterative use of MLLR is investigated in the context of large-vocabulary speaker-independent transcription of both noise-free and noisy data. It is shown that iterative application of MLLR can be beneficial especially in situations of severe mismatch. When word lattices are used, it is important that the lattices contain the correct transcription, and it is shown that global MLLR based on rough initial transcriptions of the data can be very useful in generating high-quality lattices. MLLR can also be used in an iterative fashion to refine the transcriptions of the test data and to adapt models based on the current transcriptions. These techniques were used by the HTK large-vocabulary speech recognition system for the November 1995 ARPA H3 evaluation. It is shown that iterative-application MLLR proved to be very effective prior to lattice generation and for iterative refinement.