A voice activity detection algorithm in spectro-temporal domain using sparse representation

This paper describes a new algorithm for voice activity detection (VAD), based on sparse representation of spectro-temporal domain. Our audio classification algorithm is based on multi-scale spectro-temporal modulation features which are extracted using auditory cortex model. The key concept in sparse representation is that any speech fragment can be represented as a linear combination of a small number of exemplar speech tokens. In this algorithm, the approach transforms the speech into spectro-temporal domain resulting in its decomposition into auditory-based features with multiple scales of temporal and spectral resolutions; in the next stage, each frame is divided into several sub-cubes in the new domain; then the algorithm detects the speech in the signal by using the sparse representation of sub-cubes of the frames in this domain. Simulation results are given to illustrate the effectiveness of our new VAD algorithms. The results reveal that the achieved performance is 90.11 and 91.75% under − 5 db SNR in white and car noise respectively, outperforming most of the state of the art VAD algorithms.