Convolutive Independent Component Analysis for Blind Separation of Real Multichannel Recordings of Speeches and Musics

• Main Authors: Sih-Huei Chen, 陳思卉
• Other Authors: Jiann-Ming Wu
• Format: Others
• Published: 2014
• Online Access: http://ndltd.ncl.edu.tw/handle/85m7g3

碩士 === 國立東華大學 === 應用數學系 === 102 === This work proposes a novel approach for convolutive independent component analysis with applications to blind separation of real multi-channel recordings of speeches and musics. Multichannel recordings are modeled as linear convolutive mixtures of independent sources. The proposed approach can be described as a multilayer recurrent network. Each layer performs sequential noise cancellation and blind deconvolution for estimating the dominant source of a specific recording. Empirical studies include simulation of the convolutive ICA model, estimation of the filter length of real recordings, blind separation of simulated multichannel recordings and real multichannel recordings. Quantitive performance evaluation show encouraging results.