Application of efficient score function estimation in blind speech-music separation

In this paper speech-music separation using blind source separation is discussed. The separating algorithm is in the time domain and based on the mutual information minimization. Also the natural gradient algorithm is used for its minimization. In order to do that, score function estimation from observation signal samples is needed. The accuracy and the speed of the mentioned estimation will affect on the quality of the separated signals and the processing time of the algorithm. The score function estimation in the presented algorithm is based on the fast FFT-based kernel density estimation method. The experimental results of the presented algorithm on the speech-music separation and comparing to the separating algorithm which is based on the minimum mean square error estimator, indicate that it can cause better performance and less processing time than other methods.