Voice Activity Detection Based on Discriminative Weight Training Incorporating an Output Feedback Approach

초록

In this paper, we apply an output feedback approach to the minimum classification error (MCE) method for statistical model-based voice activity detection (VAD). To exploit the inter-frame correlation of voice activity efficiently, we propose a novel technique to incorporate the decision statistic of the previous frame into the input feature vector of the MCE technique. The proposed decision statistic is expressed as the arithmetic mean of the optimally weighted features including both the likelihood ratios (LRs) of the current frame and the previous VAD decision statistic. Experimental results show that the VAD based on the MCE method incorporating the output feedback technique outperforms the VAD based on the conventional MCE method under various conditions.