A technique for noise robust voice activity detection under uncontrolled environment

Voice activity detection (VAD) is a critical component in speech processing systems. Traditional VAD methods work well in clean and controlled environments but perform poorly in real-life scenarios where various noise sources interfere with the speech signal. This article presents a technique for achieving noise robust VAD under such adverse conditions. The proposed system consists of a background noise suppression module based on the minimum mean square error spectrum power estimator using zero crossing (MMSE-SPZC), which is added before vector quantization-based VAD (VQ-VAD). Through extensive experimentation on the NOIZEUS and NIST-SRE10 databases with varying levels of noise, the effectiveness of the proposed technique is demonstrated. The results indicate substantial improvements in VAD accuracy, even in the presence of background noise. We provide an open-source implementation of the method. https://sites.google.com/view/thimmarajayadavag/downloads.