A Computation Efficient Voice Activity Detector for Low Signal-to-Noise Ratio in Hearing Aids
Fangqi Liu, Andreas Demosthenous · 2021
This paper proposes a spectral entropy-based voice activity detection method, which is computationally efficient for hearing aids. The method is highly accurate at low SNR levels by using the spectral entropy which is more robust against changes of the noise power. Compared with the traditional fast Fourier transform based spectral entropy approaches, the proposed method of calculating the spectral entropy using the outputs of a hearing aid filter-bank significantly reduces the computational complexity. The performance of the proposed method was evaluated and compared with two other computationally efficient methods. At negative SNR levels, the proposed method has an accuracy of more than 5% higher than the power-based method with the number of floating-point operations only about 1/100 of that of the statistical model based method.