Using the conformal embedding analysis to compensate the channel effect in the i-vector based speaker verification system

Abstract

The I-vector approach to speaker recognition has become the prevalent paradigm over the past 2 years, showing top performance in NIST evaluations. This success is due mainly to the capability of the I-vector to capture and compress the speaker characteristics at low dimension and the subsequent channel compensation techniques that minimize channel variability. The Linear Discriminative Analysis (LDA) followed by Within-Class Covariance Normalization (WCCN ) and Cosine Similarity Scoring (CSS) represents the best compromise between performance and computational complexity. In this paper, we propose to use Conformal Embedding Analysis (CEA ); a recently proposed manifold leaning technique; to tackle the main limitations of LDA which are: the Gaussian assumption on the classes distribution, the inability to preserve the local geometric relationships of the data-space and its reliance on the Euclidean distance for characterizing the relationships between feature vectors. Experimental results on the challenging MOBIO-voice database show that CEA+WCCN outperforms LDA+WCCN for both male and female speakers at all operating points.