Negentropy based voice-activity detection for noise estimation in very low SNR condition

Abstract

The spectral organization in a noisy speech is different during noise-only portions and speech portions. Such chaotic difference can be tracked by using negentropy as a measure. Accordingly, a Voice Activity Detector (VAD) based on the negentropy measure of magnitude of spectral components of the pseudo-stationary segments of speech is proposed and has been used for the estimation of noise statistics from the noisy speech signal in very low SNR conditions. The results show the effectiveness and usefulness of the proposed method under very noisy conditions.