Voice activity detection using periodioc/aperiodic coherence features

This paper introduces novel features for Voice Activity Detection (VAD). They are based on the coherence function between the considered frame and its LPC residue, calculated for both periodic and aperiodic components. The development of these features was motivated by the possible distinction between the periodicity and the aperiodicity character of speech and noise frames. Two statistical based decision techniques are used, they are the Discriminant Analysis (DA) and Gaussian Mixture Models (GMM) based bayesian classifier. We tested the proposed VAD technique on TIMIT database. We obtain consistent improvement as compared to features without periodic and aperiodic decomposition. In addition, we obtain encouraging results in real environmental noise.