Musical-Noise-Free Blind Speech Extraction Using ICA-Based Noise Estimation with Channel Selection

Abstract:
In this paper, we propose a modified musical-noise-free blind spatial subtraction array (BSSA) based on ICA-based iterative noise estimation with channel selection. In our previous study, we have proposed a modified BSSA consisting of dynamic noise estimation by ICA and musical-noise-free iterative spectral subtraction (SS), where multiple iterative SS are applied to each of channels. This method achieves less musical noise property, but instead always suffers from large speech distortion because of no justification of applying ICA to such signals nonlinearly distorted by SS. In this paper, first, we theoretically clarify that the degradation in ICA obeys an amplitude variation in room transfer functions between the target user and microphones. Next, we introduce a channel selection strategy into ICA, where we automatically choose less varied inputs to maintain high accuracy of noise estimation. From objective and subjective evaluations, we reveal that the proposed method outperforms the conventional method. Index Terms — Blind speech extraction, musical-noise-free, higher-order statistics, channel selection