Improvement of Speaker Vector-Based Speaker Verification

Abstract

This paper describes the improvement in the performance of a text-independent speaker verification based on a speaker vector. The verification system is based on the technique of anchor models. In our previous work, the performance improvement could be obtained by using phonetic-based models instead of Gaussian mixture models (GMMs) in speaker identification. This is because the phonetic models can represent a detailed difference in pronunciation. Therefore, we aim to improve the performance of speaker verification by using phonetic-based modeling. Comparative experiments between GMMs and Hidden Markov Models (HMMs) were conducted in the speaker verification task. In the experiments, the EER of 2.68% was obtained at 1000-dimensional speaker space when HMMs were used as anchor models.