Robust voice activity detection using selectively energy features
Junichiro Wakasugi, Noboru Hayasaka, Youji Iiguni · 2014
In this paper, we propose a robust voice activity detection algorithm that can switch the calculation method automatically depending on the noise in order to adapt various noise. We use entropy as an indicator for judging whether the noise is narrow-band or wide-band. Under narrow-band noise condition, spectral product is the suitable calculation method, on the other hand, under wide-band noise condition, using spectral summation is the suitable one. The proposed method decides the type of noise by entropy, then uses the suitable calculation method depending on the noise. We evaluated the proposed method compared with other conventional methods by ROC curves and the number of correct-segments. As the result of the experiments, the proposed method can detect the speech-segments more correctly than the other methods and shows the better performance in frame-level. The experimental result shows the proposed method can switch the calculation method appropriately depending on the noise.