A Subspace Approach for Speech Enhancement Using Frame-Level AdaBoost Classification

An efficient approach for noise suppression in speech is presented in this paper. The basic idea is taken from signal subspace analysis, which is employed for noise reduction in speech. Different other subspace approaches assumed the estimation of noise from the initial silence or unvoiced portion of the speech signal and the noise was considered to be same and present throughout the signal. The problem of noise suppression and its solution has to be modified when the noise becomes time-varying and has to be detected more accurately, instead of its crude estimation from unvoiced speech or silence. This needs an accurate classification for noise of different types which are experienced practically. In this paper an AdaBoost classifier is used to discriminate noise (colored/white, persistent/time-varying) from the corrupted voiced part of the speech. In this way the noise can be detected and estimated in each noise-only frame. Then the signal subspace technique can be used for speech enhancement. This approach produces significantly better result as compared to other signal subspace speech enhancement techniques.