Convolutive Blind Source Separation for mixed audio signals using lifting schemes

Abstract

This paper describes an efficient level decomposition for convolutive Blind Source Separation (BSS) for mixed audio signals using lifting wavelets. The convolutive mixed audio signals consist of source signals and same amount of delay or echo of the same source signal. To achieve BSS for convolutive mixtures, wavelet based Independent Component Analysis (ICA) is considered. First mixture is convolutive mix of Speech and Music. The second mixture is convolutive mix of Speech and speech. The mixed signals were decomposed by lifting wavelet from level 4 to 7. ICA is applied on each level signals then the separated signal is reconstructed using inverse lifting wavelet. The experimental results show that mixtures 2 produce better results than the mixture 1. As the decomposition level increases the Signal to Noise Ratio (SNR) also increases. The best SNR obtained for mixture 1 is 20.8261 dB in 7th level. The best SNR obtained for mixtures 2 is 21.2019 dB in the 7th level.