Combining Beamforming and Blind Source Separation to Improve Source Separation Performance

Abstract: Beamforming (BF) and Blind Source Separation (BSS) are always two interesting methodologies to witness in order to separate two sources. BSS in frequency domain have been facing a serious issue of permutation ambiguity while performing source separation using Independent Component Analysis (ICA). Permutation ambiguity is a problem of mismatch of any frequency lines between the sources, so the separation in the time domain cannot exhibit a perfect separation due to the frequency components of other sources present in the time signal of one source. Various methods have been adopted all through the years of research to get rid of this critical issue and no perfect results are produced so far. Beamforming is done with spherical waves where the array is designed according to the corresponding chosen frequency band. The distance between the microphones is always set to be less than half of the source wavelength just in order to avoid aliasing issue. When BF is designed with good resolution, using this beamforming information will give a better insight to work with BSS. The proposed method of combining BF to BSS seems to be a good approach as BF mainly depends on time-difference of arrival information (delay) between the reference microphone to the consecutive microphones. The original delay information is compared to the estimated delays for each frequency lines in order to realign the frequency lines if they see a permutation by ICA. So, there is no possibility of frequency is match still existing when the delay information is operated as a major concern. This method is compared with a method called envelope continuity which uses correlation approach between neighboring frequency lines to prove their spectral continuity. This method is also one of the wisest approaches to solve the problems of frequency domain BSS. But failure in aligning one particular frequency may lead to failure of all other successive bins which cannot be avoided. So, BF delay information is used over envelope continuity to separate sources effectively. The performance is measured using Signal to Interference Ratio measurement where Beamforming approach seems to have an improved performance compared to envelope continuity. Simulation results show performance comparison. The algorithm is tested using two speech sources in a non-reverberant environment and the sources are filtered only by delays. We use Short Time Fourier Transform (STFT) for frequency domain transformation. The Then, performance is tested in four different environments and scenarios-changing source spacing, changing microphone spacing, changing the height of source plane from microphone plane and changing the filter lengths of STFT.