Improved voice activity detection of noisy speech
Min-Sung Koh, Margaret Mortz · The Journal of the Acoustical Society of America · 2000
This paper reports on the performance of a new algorithm for detecting voice activity in the presence of additive Gaussian noise. The proposed algorithm enhances the noisy speech before determining voice activity. In the speech enhancement, two noise reduction stages are used. The first stage uses Donoho’s wavelet-based noise-reduction method. The second stage uses nonlinear spectral magnitude enhancement in the Fourier domain. After signal conversion back to the time domain and renormalization, simple energy thresholds are used to detect the presence of voice activity. Experimental results show that the new method can achieve good performance with preprocessing a speech signal even though a simple energy-based voice-activity detector is used. A new algorithm shows significantly better performance when it is compared with the voice-activity detector specified in the ITU standard G.729 Annex B and other methods proposed recently.