Parametric-Pearson-based independent component analysis for frequency-domain blind speech separation

Separation performance is improved in frequency-domain blind source separation (BSS) of speech with independent component analysis (ICA) by applying a parametric Pearson distribution system. ICA adaptation rules include a score function determined by approximated source distribution, and better approximation improves separation performance. Previously, conventional hyperbolic tangent (tanh) or generalized Gaussian distribution (GGD) was uniformly applied to the score function for all frequency bins, despite the fact that a wideband speech signal has different distributions at different frequencies. To obtain better score functions, we propose the integration of a parametric Pearson distribution system with ICA learning rules. The score function is estimated by using appropriate Pearson distribution parameters for each frequency bin. We consider three estimation methods with Pearson distribution parameters and conduct separation experiments with real speech signals convolved with actual room impulse responses. Consequently, the signal-to-interference ratio (SIR) of the proposed methods significantly improve over 3 dB compared to conventional methods.