Robust voice activity detection for DTX operation of speech coders

Robust detection of voice activity for short-term speech frames is essential for discontinuous transmission (DTX) mode of operation of vocoders such as IS-641. A reference VAD for the IS-641 coder has been chosen for such a purpose and is based on the GSM-EFR (enhance full rate) VAD. We show by developing a comprehensive evaluation procedure that the reference VAD is sensitive to speech level variations. For example, a significant increase is seen in frames falsely classified as active at speech levels of 10 dB above or below nominal level. We propose a solution based on automatic gain control to reduce level sensitivity. Objective performance measures confirm the robustness of our proposed VAD.