A Robust Feature Extraction for End Point Detection in the Nonstationary Noisy Environment

This paper proposes a robust feature extraction for voice activity detection (VAD) in noisy environments where nonstationary noises exist. The accuracy of the VAD is drastically reduced because the fluctuation of features at the noise intervals causes false alarm rate to being increased. In this paper, in order to improve the VAD accuracy, harmonic-weighted energy is proposed. This feature extraction method focuses on voiced speech interval and weights the amount of the ‘harmonicity’ to the averaged energy of the frame input. To evaluate the performance of the proposed feature extraction method, receiver operating characteristic curves and equal error rate are measured. From the results, it is proved that the proposed method is the discriminative feature for VAD. Keywords— voice activity detection, robust feature extraction, harmonic to noise ratio