Özet:
Blind Source Separation (BSS) is one of the most challenging problems in the field of audio and speech processing. Many different methods have been proposed to solve BSS problem in the literature. In addition, speaker recognition systems have gained considerable interest from researchers for decades due to the breadth of their field of application.
In this study, we have compared the performance of three popular BSS methods implementations: Fast-ICA, Kernel-ICA and Fast-IVA which are based on Independent Component analysis (ICA) and Independent Vector Analysis (IVA) respectively. Initially, classical performance comparison metrics such as Source-to-Artifact Ratio, Source-to- Distortion Ratio, Source-to-Noise Ratio, are implemented for comparison.
For further investigation, speaker recognition system has been developed to examine the effect of speech separation on the performance of these recognition systems.
In our experiments, we used two data set the first one is in Arabic languge and contains voice records frome 13 speaker: 3 female , 10 male.the second data set is the ELSDSR data which in English languge and contains voice records from 22 speakers: 10 female, 12 male.
The performance of BSS methods is measured under four scenarios. The first three is composed to see the effect of noise. Therefore, we used the mixture of clean source signals, the mixture of source signals with additive Gaussian noise, adding Gaussian
noise to clean source mixture. In the fourth scenario, we applied speaker recognition system based on Gaussian mixture models (GMMs) and I-vectors, the performance of the speaker recognition system is measured by Equal Error Ratio (EER), which is, the most reliable measurement in this field.
Experimental results show that the Fast-IVA has better performance than the Fast-ICA method according to performance metrics used in this study. In terms of EER, I-vector gives the better result than GMM for separated signals by IVA and ICA.
