Welcome to the companion page for the paper “Music Source Separation via Stem Discovery”! Here you will find audio examples of separated stems for the MoisesDB test set.
Sound player provided by trackswitch.
The player offers two levels of selection: the first level picks what you want to hear — the full Mix or a single stem such as Vocals or Drums. The second level toggles between the systems producing that audio: the reference Ground Truth stems from MoisesDB, the Autoencoder reconstructions, the SAM-Audio separations, or the Mus3D stems estimated by our model. The mix sums the stems of the selected system.
Examples are summed to mono since the separation model output is mono.
These examples demonstrate the benefit of audio queries in separation (Mus3D) compared with text-prompts (SAM-Audio) and the artifacts induced by the autoencoder part of the models.
Since we do blind separation (Mus3D), the stems sometimes include instruments that are not in the reference audio files, for example a-tonal percussion, synth pad and toms in the third example.