Extracting a singing voice from its music accompaniment can significantly facilitate certain applications of Music Information Retrieval including singer identification and singing melody extraction. In this paper, we present a hybrid approach for this purpose, which combines properties of the Azimuth Discrimination and Resynthesis (ADRess) method with Independent Component Analysis (ICA). Our proposed approach is developed specifically for the case of singing voice separation from stereophonic recordings. The paper presents the characteristics of the proposed method and details an objective evaluation of its effectiveness.
Stratis Sofianos, Aladdin M. Ariyaeeinia, Richard