Symmetric Distortion Measure for Speaker Recognition
Evgeny A. Karpov, Tomi Kinnunen · 2004
Abstract We consider matching functions in vector quantization (VQ)based speaker recognition systems. In VQ-based systems, aspeaker model consists of a small collection of representativevectors, and matching is performed by computing a dissim-ilarity value between the unknown speaker’s feature vectorsand the speaker models. Typically, the average/total quanti-zation error is used as the dissimilarity measure. However,this measure lack the symmetricity requirement of a properdistance measure. This is counterintuitive because matchscore between speakers X and Y is different from the matchscore between Y and X . Furthermore, the distortion measurecan yield a zero value (perfect match) for non-identical vec-tor sets, which is undesirable. In this study, we study ways ofmaking the quantization distortion functions proper distancemeasures. The study includes discussion of the theoreticalproperties of different measures, as well as an evaluation ona subset of the NIST99 speaker recognition evaluation cor-pus.