Authors : Mostafa Esmaeilbeig, Hamid Sheikhzadeh and Farbod Razzazi
Abstract: In this study, we propose a novel algorithm for solving the permutation ambiguity problem in convolutive blind source separation of speech signals. Transferring convolutive mixtures into time-frequency domain, enables us to separate source signals by employing instantaneous algorithms in each frequency bin. After separation, the main challenge is the scale and permutation ambiguities which can imperil the separation performance. Overcoming this challenge needs the reordering of all separated signals in each frequency bin according to order of source signal. In this study we propose a new algorithm for reordering the separated signals based on statistics of MFCC of speech signals. In each frequency bin, the separated subband signals are transferred back to time-domain and their individual MFCC’s are extracted. Then, based on simple statistics of the MFCC’s the permutation problem is resolved. The proposed algorithm drastically decreasesthe computational complexity and as a result speeds up the permutation correction process.
Mostafa Esmaeilbeig, Hamid Sheikhzadeh and Farbod Razzazi, 2016. Permutation Correction in Convolutive BSS Using MFCC Coefficients. International Journal of Signal System Control and Engineering Application, 9: 157-165.