Improving single frequency filtering based Voice Activity Detection (VAD) using spectral subtraction based noise cancellation

Voice Activity Detection (VAD) is a basic front end step in all of the speech processing engines. There have been proposed many Time domain and Frequency domain algorithms with different computational complexity. Common artifacts in VAD are Front End Clipping (FEC), Mid Speech Clipping (MSC), over clipping and Noise detected as Speech (NDS). Performance of VAD is dependent on SNR between speech and background noise. Background noise could be either a wide band noise spanning entire speech frequency band or it might be a narrowband noise which does not interfere with the speech frequency band. Simple VAD methods based on short time energy and zero crossing detection fails to discriminate between speech frame and noise frame in low SNR environment. The frequency domain methods have higher accuracy in low SNR environments. In Single Frequency Filtering (SFF) approach instead of computing the complete FFT of the given audio frame, the power envelope of the spectrum at a discrete frequency interval 20Hz in the speech band of 20Hz to 4KHz are computed. The accuracy of the SFF VAD can be improved by applying an adaptive Noise Canceller prior to SFF VAD. The VAD decision of SFF approach at the transition from speech to noise and vice versa is improved to a remarkable extent by using an adaptive noise cancellation. Evaluation of SFF approach was done with Adaptive Spectral Subtraction based Noise Cancellation prior to SFF VAD.