Robust voice-activity detection based on the wavelet transform
Joachim Stegmann, G. Schroder · 2002
This paper describes a new approach to voice-activity detection (VAD) which is based on the wavelet transform (WT). The algorithm utilizes the WTs flexibility in the time-frequency resolution to compute robust parameters for VAD decision. Furthermore, it exhibits a low complexity and can be easily adapted to operate as a pre-processor for many speech-coding algorithms. Two versions of the wavelet-transform-based VAD (WT-VAD) are tested against the VAD of the ITU-T G.729 Annex B (G729) and the VAD of the GSM enhanced full-rate codec (GSM), respectively. For a variety of background-noise types the WT-VAD shows superior noise robustness to signal-to-noise ratios above 10 dB.