Multi frequency resolution method for solving scaling problem in frequency domain blind source separation

Abstract

Blind Source Separation (BSS) is the method for separating source signals from the mixture of sources. For BSS, Independent Component Analysis (ICA) that assumes the statistical independence among sources is widely used. In reverberant condition, frequency domain ICA can be used for BSS, as the convolu-tion in time domain can be changed into the simple multiplication in frequency domain. However it accompanies permutation and scaling problem that result in the distortion of separated outputs. Permutation problem is the ambiguity in the order of output at each frequency bin, and scaling problem is scaling ambiguity among frequency bins. There are existing methods for solving permutation and scaling problem. For instance, independent vector analysis has been proposed for permutation problem, and minimal distortion principle has been proposed for scaling problem. Multi frequency resolution method is proposed for solving scaling problem by comparing two separated signals with different resolution in frequency domain. To solve scaling problem, estimate the ratio between adjacent frequency bins by solving quadratic problem and rescale bins to fit the scale between them. In experiment, our proposed method is applied to BSS for 2 sources firstly for IVA by varying parameters. MFR showed improvement of the performance judging from increased cross correlation with source signal in frequency domain. Hybrid method that adds MDP and MFR was proposed in this thesis, but it did not show much improvement. Multi frequency resolution method for solving permutation problem is proposed in addition. It calculates cost function for all permutation candidates with the method used for solving scaling problem. Then align the order of outputs followed by the case that has the minimum cost function.