Voice Activity Detection based on Inverse Normalized Noise Likelihood Estimation
Tomas Dekens, Mike Demol, Werner Verhelst, Frédéric Beaugendre · 2007
Abstract—In this paper we develop a voice activity detection algorithm based on the likelihood that only noise is present in the current signal frame. For this we exploit the fact that the Fourier coefficients of most noise processes can be modeled as statistically independent Gaussian random variables. We also give an overview of different voice activity detectors previously described in the literature and compare their results to the ones obtained with the voice activity detector we propose in this paper. According to our tests, at high speech detection probabilities, the proposed algorithm shows results than are comparable to or better than the other voice activity detectors we consider, while the simplicity of the algorithm ensures low computational complexity. Key Words—Noise estimation, Speech enhancement, Voice activity detection.