Analysis and improvement of a statistical model-based voice activity detector

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