Presentation Information
[A-8-18]Performance Evaluation of Orthogonalization Methods for Audio Source Separation Using FastICA
◎△Shoki Shimizu1, Yusuke Endo1, Koujin Takeda1 (1. Graduate School of science and Engineering. Ibaraki Univ.)
Keywords:
ICA,Audio Source Separation,FastICA,Amari Distance,Negentropy
Independent Component Analysis (ICA) is used in fields like biomedical signal processing and telecommunications, and widely applied in audio source separation. The FastICA algorithm is known for its fast convergence. When separating multiple sources, an orthogonalization constraint is essential to prevent estimated vectors from converging to the same source. This paper evaluates two orthogonalization methods—deflationary and symmetric—on separation performance via simulations.Varying the source count from 5 to 50, we compared both methods using negentropy (indicating independence) and Amari distance (measuring estimation accuracy). As the source count increased, deflationary orthogonalization degraded in performance due to cumulative errors from repeated projections. In contrast, symmetric orthogonalization maintained stable accuracy. Hypothesis testing confirmed significant differences for negentropy and Amari distance. These findings demonstrate that symmetric orthogonalization, by processing all components simultaneously, prevents error bias and achieves stable, high-precision separation in multi-source environments.