Publications
カンファレンス (国内) A Generalized Minimal Distortion Principle to Solve the Scale Ambiguity in Blind Source Separation
シャイブラー ロビン
日本音響学会 2020年秋季研究発表会 (ASJ 2020 autumn)
2020.9.9
This paper addresses the problem of blind source separation (BSS), where the goal is to separate mixed audio signals into independent sources without prior information. One challenge is the ambiguity due to arbitrary scaling of sources. This study focuses on estimating source images, representing source signals as perceived at microphone locations. Existing methods include projection back (PB) using the inverse demixing matrix and minimal distortion principle (MDP) respecting input microphone signals. The paper introduces a novel approach by extending MDP to maximum likelihood estimation (MLE) using a mixed-norm model for residual spectrograms, enabling fine sparsity control. The proposed method employs majorization-minimization (MM) for optimization, resulting in an iteratively reweighted least-squares technique. Evaluation demonstrates potential separation improvement of up to 2 dB with parameter tuning, maintaining low distortion and computational cost, offering an effective enhancement for blind source separation methods.