Dual Microphone Voice Activity Detection Exploiting Interchannel Time and Level Differences

Abstract

The two most important spatial cues in human auditory system may be the interaural time difference and the interaural level difference. There have been many attempts to utilize the time difference of arrival (TDoA) and level difference between two microphone signals for voice activity detection (VAD). In this letter, we propose a dual microphone VAD algorithm based on a support vector machine for which the input vector consists of both TDoA-based and level difference-based features. Several candidates for the feature combination have been compared using various TDoA-related and level difference-related features. Experimental results showed that the proposed VAD algorithm outperformed a standardized single microphone VAD, VADs based on the TDoA or level difference, and logical combination of them in various noise environments. © 2016 IEEE.