A non-causal approach to voice activity detection in adverse environments using a novel noise estimator

Analyzing the characteristics of the LR-based VAD, it was found that the delay associated with the decision directed (DD) a priori SNR estimator can lead to detection errors at speech onsets and offsets. In this paper the properties of a non-causal estimator, used before in a speech enhancement system, are investigated. It is shown that the application of the non-causal estimator improves the robustness of the VAD in noisy environments, specifically at low SNRs. In addition, the associated noise estimation procedure has been further improved by the application of a dynamic time varying smoothing factor. Objective tests conducted based on speech/non-speech discrimination show that the proposed VAD outperforms standard VAD algorithms, including ETSI-VADNest, AMR1, AMR2, and also the statistical VADs based on smoothed LR and multiple observation LR, specifically at low SNRs, at the cost of some delay.