Towards shifted NMF for improved monaural separation

Abstract

The ability of Non-negative Matrix Factorisation (NMF) to decompose magnitude spectrogram into meaningful entities has found use in many audio applications. NMF can be used to factorise audio spectrogram of a music signal into parts based frequency basis functions which typically corresponds to notes and chords in music. However, these pitched basis functions needed to be clustered to their respective sources. Many clustering algorithms have been proposed to group these basis functions. Recently, Shifted Non-negative Matrix Factorisation (SNMF) based methods have been used to reconstruct individual sound sources. Clustering of basis functions using SNMF uses a Constant Q Transform (CQT) of the frequency basis functions. Here, we argue that incorporating the CQT into the SNMF model can be used to better the separation quality of individual sources. An algorithm is presented to estimate sound sources and is an improvement to the existing techniques. Results are compared to show the improvement.