A Mahalanobis Distance Scoring with KISS Metric Learning Algorithm for Speaker Recognition

摘要

The cosine similarity scoring is often used in the i-vector model for its computational efficiency and performance in text-independent speaker recognition field. We propose a new Mahalanobis distance scoring with distance metric learning algorithm in this paper. The Mahalanobis metric matrix is learned using the KISS (keep it simple and straightforward!) method, which is motivated by a statistical inference perspective based on a likelihood-ratio test. After whitening and length-normalization, the i-vectors extracted from the development utterances were used to train the metric matrix. Then, the score between the target i-vector and the test i-vector is based on the Mahalanobis distance. The results on NIST 2008 telephone data show that the performance of new scoring is obviously better than the cosine similarity scoring's.