Multi-class UBM-based MLLR m-vector system for speaker verification

In this paper, we extend the recently introduced Maximum Likelihood Linear Regression (MLLR) super-vector based m-vector speaker verification system to a multi-class MLLR m-vector system. In the conventional case, global class MLLR transformation is estimated with respect to Universal Background Model (UBM) for a given speech data, which is then used in the form of super-vector for m-vector system. In the proposed system, Gaussian mean vectors of the UBM are first clustered into several classes. Then, MLLR transformations are estimated (of a speech data) for each class, and are used in the form of super-vectors for speaker characterization using the m-vector technique. We consider two clustering approaches: one is based on the conventional K-means and the other is proposed based on Expectation Maximization (EM) and Maximum Likelihood (ML). Both systems yield better performance than the conventional m-vector system and allow for multiple MLLR transforms without additional temporal alignment of the data with respect to UBM. Furthermore, we show that, contrary to conventional K-means, the proposed clustering is not affected by the random initialization, and also provides equal or comparable system performance. The system performances are shown on NIST 2008 SRE core condition over various tasks.