Complex ISNMF: a phase-aware model for monaural audio source separation

Paul Magron and Tuomas Virtanen.

IEEE/ACM Transactions on Audio, Speech, and Language Processing, Vol.27, no. 1, pp. 20-31, January 2019

On this page, we provide several audio excerpts that illustrate the source separation experiments presented in the paper. The songs are extracted from the Demixing Secrets Database (DSD100).

We notice that all Wiener filters (Wiener, AW and CAW) yield approximately similar results with a strong musical noise. This confirms that those methods usually perform well in Oracle or strongly informed settings, as shown in some of our previous studies.
The CNMF technique tends to partly remove the musical noise, though it introduces some smearing artifacts and interference (notably between the "vocals" and "other" tracks).
The improvement of the proposed Complex ISNMF method is mainly noticable in the bass track: it reduces the musical noise artifacts compared to the Wiener filters, and retrieves bass signals with slightly neater attacks.