Lattice MLLR based m-vector system for speaker verification

The recently introduced m-vector approach uses Maximum Likelihood Linear Regression (MLLR) super-vectors for speaker verification, where MLLR super-vectors are estimated with respect to a Universal Background Model (UBM) without any transcription of speech segments and speaker m-vectors are obtained by uniform segmentation of their MLLR super-vectors. Hence, this approach does not exploit the phonetic content of the speech segments. In this paper, we propose the integration of an Automatic Speech Recognition (ASR) based multi-class MLLR transformation into the m-vector system. We consider two variants, with MLLR transformations computed either on the 1-best (hypothesis) or on the lattice word transcriptions. The former case is able to account for the risk of ASR transcription errors. We show that the proposed systems outperform the conventional method over various tasks of the NIST SRE 2008 core condition.