In this paper, voice activity detection (VAD) is posed as a binary detection problem of an unknown speech signal in both stationary (vehicular) and nonstationary (babble) noise environments. Optimal detection methods are applied on the wavelet transform coefficients of a signal segment to determine the presence of speech. Theoretical analysis is done to justify the effectiveness of multiresolution decomposition on the computation of the noise eigenvalues and vectors and sufficient statistics. VAD results are compared to optimal detection without wavelet transformation and to an energy based method which is used as control. The results show the superiority of the proposed method as increased accuracy in detection.
展开▼
