Using LSF features for speaker verification in noise

Abstract

An automatic, text-independent speaker verification (SV) system is proposed using Line Spectral Frequency (LSF) features. The state-of-the-art Gaussian Mixture Model with Universal Background Model (GMM-UBM) framework is used for speaker modeling and verification. A score-level fusion based technique is employed in order to extract complementary information from static and dynamic LSF features and improve the noise-robustness of the SV system. In addition, the speaker-discriminative power of different speech zones such as vowels, non-vowels, and transitions are investigated. Rapidly varying transition regions of speech are found to be most speaker-discriminative in high SNR conditions. Steady, high-energy vowel regions are robust against noise and are most speaker-discriminative in low SNR conditions. We show that selectively utilizing features from a combination of transition and steady vowel zones further improves the performance of the score-level fusion based SV system under noisy conditions.