학술논문

Convolutive transfer function-based independent component analysis for overdetermined blind source separation
Document Type
Conference
Source
2022 16th IEEE International Conference on Signal Processing (ICSP) Signal Processing (ICSP), 2022 16th IEEE International Conference on. 1:22-26 Oct, 2022
Subject
Signal Processing and Analysis
Wiener filters
Transfer functions
Independent component analysis
Signal processing
Linear programming
Blind source separation
Reverberation
convolutive transfer function
independent component analysis
Language
ISSN
2164-5221
Abstract
Frequency domain independent component analysis (FDICA) is a fundamental and widely used method for blind source separation. However, the performance of FDICA degrades in highly reverberant environments because of the limitation of the narrowband assumption. This paper proposes a convolutive transfer function (CTF) based independent component analysis. Compared to the narrowband assumption, the CTF approximation results in fewer errors to represent the time-domain convolutive mixture with long reverberation times, and hence achieves improved separation performance. Moreover, the CTF enables the use of a short frame size to represent long room impulse responses, which increases the accuracy of statistical estimation. We formulate the objective function within the maximum likelihood framework. The optimization of the objective function is accomplished by designing an auxiliary function based on the majorization-minimization principle in the overdetermined case. Finally, the scale-fixed source is recovered by filtering the mixture through a multichannel Wiener filter. In addition, we show that FDICA is a special case of the proposed framework. Experimental results show the efficacy of the proposed method.