A Novel Self-adaptive Voice Activity Detector Based on Robust Logistic Regression

Voice Activity Detection (VAD) is a fundamental preprocessing component in advanced speaker verification systems. Robustness against noise and real-time operation are two essential characteristics of a practical voice activity detector. In this paper, we propose to use a probabilistic framework based on Robust Logistic Regression (RLR) as a discriminative variant of the self-adaptive-Vector Quantization VAD (VQ-VAD). Robust logistic regression overcomes the intrinsic limitation of regular logistic regression in case of complete class separation. The performance of the proposed VAD for the task of speaker verification is evaluated using the clean TIMIT database and its noise-corrupted versions. The conducted experiments using the GMM-UBM baseline for speaker verification demonstrate that the proposed VAD can provide an efficient way to select high-quality speech frames in noisy environments with global convergence property and computational efficiency compared to VQ-VAD. Interestingly, the performance of the speaker verification system based on the proposed VAD in terms of Equal Error Rate EER compares favorably to the VQ-VAD based system, especially at lower signal to noise ratio.