Voice Activity Detection Algorithm Based on Mel-scale Frequency Log-Spectral Energy Difference In Noise Environment

Voice Activity Detection (VAD) is an important part of the speech signal processing, its accuracy directly influences the speed and result of the speech signal processing. Most means of VAD is done in laboratory-scale environment, it requires stationary noise and high Signal Noise Ratio (SNR). But in fact, these conditions above can´t be satisfied usually. A VAD algorithm is put forward based on “Mel-scale Frequency Log-Spectral Energy Difference”, which has easy distance-measuring degree and clear physical sense. Compared with the traditional method, this algorithm used the relative dimension to replace the absolute one, so, in the case of low SNR and in the environment of slowly varying and non-stationary noise, this algorithm can demarcate speech in noise accurately, and meanwhile, it has good robustness.