Robust voice activity detection directed by noise classification

In this paper voice activity detection (VAD) is formulated as a two-class classification problem using support vector machines (SVM). The proposed method combines a noise robust speech processing feature extraction process together with SVM models trained in different background noises for speech/non-speech classification. A multi-class SVM is also used to classify background noises in order to select SVM model for VAD. The proposed VAD is tested with TIMIT data artificially distorted by different additive noise types and is compared with state-of-the-art VADs. Experimental results show that the proposed VAD can extract speech activity under poor SNR conditions, and it is also insensitive to variable levels of noise.