A maximum log-likelihood approach to voice activity detection

Modern voice activity detection (VAD) algorithms must achieve reliable operation at low signal-to-noise ratios (SNR). Although a lot of research has been performed to solve this issue, the operation of existing VAD algorithms is still far away from ideal. In this paper, we present a novel VAD algorithm, in which we apply the Teager energy cepstral coefficients, to obtain a noise robust feature extraction method, together with Gaussian mixture models that serve for the classification of speech and silence periods. In the suggested solution, the threshold method used in many noise robust VAD algorithms is eliminated, thus favoring its use in real applications. The performance of this novel algorithm was tested under known and unknown noise statistics, and compared to a statistical model-based approach found in literature. The results obtained show that the proposed solution achieves better accuracy and significantly reduces clipping of speech periods; thus achieving superior signal quality.