Distant-talking activity detection with multi-channel audio input

In this paper we focus on the problem of voice activity detectionfor distant-talking speech recognition.Under noisy conditions the exploitation ofa coherence measure between differentmicrophones turns out to be an effective featurein the case of moderate SNR andreverberant signals.A non linear processing using the phase information in the Cross PowerSpectrum leads to the detection of speech segments.The recognition test on a real multi-channel corpus ofisolated words shows that the proposed technique outperformsthe classic energy-based algorithm.