Single channel music and speech separation using non-negative matrix factorization

In this paper, non-negative matrix factorization (NMF) is used to separate speech and music signals based on a single channel recording. The assumption that if two independent zero-mean signals are added then their energies are also added has led us to develop a two-stage method (training and separation) that works on time-frequency domain. The performance of the method in separation is evaluated by observing the power of the separated signals in time-frequency domain, and by measuring the increase in signal-to-interference and signal-to-noise ratios after separation. Finally, we discuss the problems faced and the work that can be done in future to enhance the performance of the method in separation.