A voice activity detection algorithm for communication systems with dynamically varying background acoustic noise
I.D. Lee, Harold P. Stern, Sabri A. Mahmoud · 2002
Speech can be modeled as short bursts of vocal energy separated by silence gaps. During typical conversation talk-spurts comprise only 31.5% of each party's speech and the remaining 68.5% is silence. Communication systems can achieve significant gains in spectral efficiency and energy efficiency by disconnecting the users from the spectral resource during the silence periods. This paper develops a simple, efficient, and robust voice activity detection (VAD) algorithm to work in a mobile or portable environment exhibiting dynamically varying background noise. The VAD uses probabilistic distances based on the energy content, the periodicity, the stationarity and the spectral distribution within the low frequency band to decide if the presented speech frame is speech or silence.