Robust distributed multi-speaker voice activity detection using stability selection for sparse non-negative feature extraction

In this paper, we propose a robust multi-speaker voice activity detection approach for wireless acoustic sensor networks (WASN). Each node of the WASN receives a mixture of sound sources. We propose a non-negative feature extraction using stability selection that exploits the sparsity of the speech energy signals. The strongest right singular vectors serve as source-specific features for the subsequent voice activity detection (VAD). To separate active speech frames from silent frames, we propose a robust Mahalanobis classifier that is based on an M-estimator of the covariance matrix. The proposed approach can also be applied to a distributed setting, where no fusion center is available. Highly accurate VAD results are obtained in a challenging WASN of 20 nodes observing 6 sources in a reverberant environment.