Simultaneous Separation and Segmentation in Layered Music

In many pieces of music, the composer signals how individual sonic elements (samples, loops, the trumpet section) should be grouped by introducing sources or groups in a layered manner. We propose to discover and leverage the layering structure and use it for both structural segmentation and source separation. We use reconstruction error from non-negative matrix factorization (NMF) to guide structure discovery. Reconstruction error spikes at moments of significant sonic change. This guides segmentation and also lets us group basis sets for NMF. The number of sources, the types of sources, and when the sources are active are not known in advance. The only information is a specific type of layering structure. There is no separate training phase to learn a good basis set. No prior seeding of the NMF matrices is required. Unlike standard approaches to NMF there is no need for a post-processor to partition the learned basis functions by group. Source groups are learned automatically from the data. We evaluate our method on mixtures consisting of looping source groups. This separation approach outperforms a standard clustering NMF source separation approach on such mixtures. We find our segmentation approach is competitive with state-of-the-art segmentation methods on this dataset.