Independent component analysis based single channel speech enhancement

Abstract

Hands-free use of phones and voice navigation is the preferred solution in cars. However, the car environment is noisy and often-times noise substantially degrades intelligibility of speech. We propose a single channel algorithm to reduce car noise. The approach employs two phases. First, independent component analysis (ICA) is applied to a large ensemble of clean speech training frames to reveal their underlying statistically independent basis. The distribution of the ICA transformed data is estimated in the training phase. It is required for computing the covariance matrix of the ICA transformed speech data used in the operational phase. Second, a Wiener filter is applied to estimate the clean speech from the received noisy speech. The Wiener filter minimizes the mean-square error between the estimated signal and the clean speech signal in the ICA domain. An inverse transformation from ICA domain back to time domain reconstructs the enhanced signal. Extensive experiments show considerable noise reduction capability of the proposed algorithm. The evaluation is performed with respect to four objective quality measure criteria.