Distant Multi-speaker Voice Activity Detection Using Relative Energy Ratio

While single-speaker voice activity detection is a well-studied problem, multi-speaker voice activity detection (MSVAD) for distant speech recognition remains a challenging task. In this work, we propose a new MSVAD system for identifying voice activity of an individual speaker from distant speech data captured with a microphone array. In contrast to normal energy-based approaches, our MSVAD algorithm employs information from the interfering channels in a hierarchical manner in order to adaptively adjust the threshold. We demonstrate the effectiveness of our MSVAD algorithm through experiments on the Speech Separation Challenge corpus [1]. A MSVAD technique with the cross-meeting normalized energy criterion [2] provided a missed detection rate (MDR) of 7.4% with a false alarm rate (FAR) of 28.0%. By incorporating the proposed criterion in the algorithm, the MDR and FAR were further reduced to 4.3% and 19.6%, respectively. Our algorithm also achieved speech recognition performance comparable to manual segmentation results. Moreover, our method requires no parameter training and has low computational complexity.