From investigation of a statistical model-based voice activity detector (VA
D), it is found that the likelihood ratio defined in the VAD has a fundamen
tal problem at the offset regions of speech signals. Thus, we analyze the b
ehavioral mechanism of the likelihood ratio, identify the reason for the un
wanted phenomenon, and propose a solution based on a smoothed likelihood ra
tio. Objective test results show that the proposed method gives significant
improvement to the original VAD. Additionally, the improved VAD results in
performance even superior to G.729B and comparable to AMR VAD option 2.