Source Enumeration and Robust Voice Activity Detection in Wireless Acoustic Sensor Networks
Tanuj Hasija, Martin Gölz, Michael Muma, Peter J. Schreier, Abdelhak M. Zoubir · 2019
We propose a robust technique for multi-speaker voice activity detection and source enumeration in wireless acoustic sensor networks (WASN). The proposed technique first clusters the nodes that observe a single speaker as dominant source, and then estimates the voice activity of each speaker by introducing a block-sparsity penalizing term in the unmixing problem. The method is scalable in terms of the number of simultaneously active speakers, does not require setting empirical thresholds, and is robust to impulsive noise sources. The results are validated using a WASN with four human speakers and two impulsive noise sources observed by 15 nodes.