A quality measure method using Gaussian mixture models and divergence measure for speaker identification

Previous work has demonstrated the promise of frame-level quality measure methods to robust speaker recognition. This paper explores the issues involved in applying soft estimates to quality measures as weighting factors in score computation. A quality measure algorithm using Gaussian mixture density and Jensen divergence measure is presented for traditional GMMUBM scoring mechanism. Derivation and validation of the quality measurement are reported in this paper. We investigate the usefulness of different feature processing, different GMMbased quality models and incoporation of divergence measure for quality estimation. Comparison experiments performed on the NIST1999 SRE corpus show the effectiveness of the proposed method.