Study of a voice activity detector and its influence on a noise reduction system

Abstract Some noise reduction processings, such as spectral subtraction techniques, require the learning of noise characteristics. In consequence, a Voice Activity Detector (VAD) is needed to determine noise and speech sequences. In this paper, in the case of spatially uncorrelated (or slightly correlated) noises, we introduce a new technique based on the coherence function which is used to determine a speech/noise classification algorithm. We combine it with a noise reduction technique based on the spectral subtraction and evaluate its influence. We report on results obtained on the performance of the algorithm and conclude that they are quite comparable to those obtained using a manual labelling.