Singing voice separation for mono-channel music using Non-negative Matrix Factorization

As music has turned digital, much research has been shifted toward digital music processing. Singing voice separation is one of the active research areas since the singing voice itself contains abundant information within, including melody, singerpsilas characteristic, lyrics, language, emotion, etc. These wide variety of resources are quite useful for music information retrieval (MIR), singer identification, or even karaoke systems. However, this singing voice separation, especially in mono-channel environment, is a very challenging problem, whose existing methods are still impractical for real-world music. In this work, we propose an algorithm based on non-negative matrix factorization (NMF) approach to decompose spectra of music, then provide criteria for automatic component selection. Our preliminary results have demonstrated its effectiveness in pitch extraction resulting from the separated singing voice.