Endpoint Detection of Noisy Signal Based on Feature Extraction Using Independent Component Analysis

Qingbo He,Fanrang Kong,Jianping Wang,Yongbin Liu

DOI: https://doi.org/10.3969/j.issn.1004-9037.2007.01.005

2007-01-01

Abstract:The speech endpoint detection in the presence of noise is analyzed since endpoint detection methods have difficulties at low signal-to-noise ratios. This paper proposes a novel method for distinguishing the signal and the noise by the statistical property of themselves, based on higher order statistical feature extraction using independent component analysis (ICA). Because ICA transform increases the difference of the statistical property between the speech and the noise, the speech and the noise can be distinguished in the ICA domain. On the basis, the characteristics called the ICA energy (ICAE) and the filtered ICAE (FICAE) are used for the noisy signal endpoint detection. Experimental results show that the endpoint detection method combining FICAE and ICAE has certain stability at various signal-to-noise ratios, and can effectively detect the speech endpoints at very low signal-to-noise ratio. Thus the method has good robust performance and provides a new approach for detecting endpoints of weak signals under noisy environments.