VOICE ACTIVITY DETECTION IN THE DFT DOMAIN BASED ON A PARAMETRIC NOISE MODEL
Colin Breithaupt, Rainer Martin · 2006
We present a robust voice activity detection (VAD) algo-rithm which is based on the statistics of the coefficients of the discrete Fourier transform (DFT) derived from short signal segments. This algorithm uses a common para-metric noise probability density function (PDF) in all fre-quency bins. The noise model is based on a Rayleigh in-verse Gaussian distribution and adapted to the statistics of the noise during speech-absence. As only the current and past signal frames are analysed, the detection is causal and no additional delay is introduced. A framework for pro-tecting low energy syllables at the end of utterances is also described. 1.