Noise Robust Voice Activity Detection Based on Statistical Model and Parallel Non-Linear Kalman Filtering
Masakiyo Fujimoto, Kentaro Ishizuka, Hiroko Kato · 2007
This paper addresses the problem of voice activity detection in noise environments. The proposed voice activity detection technique described in this paper is based on a statistical model approach, and estimates the statistical models sequentially without a prior knowledge of noise. The crucial factor as regards the statistical model-based approach is noise parameter estimation, especially non-stationary noise. To deal with this problem, a parallel non-linear Kalman filter, that is a multiplied estimator, is used for sequential noise estimation. Also, a backward estimation is used for noise estimation and likelihood calculation for speech / non-speech discrimination. In the evaluation results, we observed that the proposed method significantly outperforms conventional methods as regards voice activity detection accuracy in noisy environments.