Introducing New Mechanism in the Learning Process of Fdica-based Speech Separation

The blind source separation for speech using frequencydomain independent component analysis(FDICA) is considered. As a source separation system, Saruwatari et al.[1] proposed a method by integrating the independent component analysis (ICA) and array signal processing. In this paper, we introduce the following two techniques into the learning process of the method[1]. (1)Classification of acquired array signals with respect to the number of speakers. (2)Direction-of-Arrival(DOA) estimation for each speaker using the intervals(frames) which are classified into singlespeaker frame. Through some experiments, we can confirm that these techniques are effective to guarantee the convergence to the global optimal solution in the learning process.