NMF with spectral and temporal continuity criteria for monaural sound source separation

Abstract

Nonnegative Matrix Factorization (NMF) is a well suited and widely used method for monaural sound source separation. It has been shown, that an additional cost term supporting temporal continuity can improve the separation quality [1]. We extend this model by adding a cost term, that penalizes large variations in the spectral dimension. We propose two different cost terms for this purpose and also propose a new cost term for temporal continuity. We evaluate these cost terms on different mixtures of samples of pitched instruments, drum sounds and other acoustical signals. Our results show, that penalizing large spectral variations can improve separation quality. The results also show, that our alternative temporal continuity cost term leads to better separation results than the temporal continuity cost term proposed in [1].