Demucs: Deep Extractor for Music Sources with extra unlabeled data remixed
This addresses source separation for music, an incremental improvement in waveform-based methods.
The paper tackles music source separation by introducing a convolutional and recurrent model that outperforms Wave-U-Net by 1.6 SDR points on waveforms, and proposes a scheme to leverage unlabeled data through remixing to create weakly supervised examples, showing waveform methods can compete with spectrogram-based approaches.
We study the problem of source separation for music using deep learning with four known sources: drums, bass, vocals and other accompaniments. State-of-the-art approaches predict soft masks over mixture spectrograms while methods working on the waveform are lagging behind as measured on the standard MusDB benchmark. Our contribution is two fold. (i) We introduce a simple convolutional and recurrent model that outperforms the state-of-the-art model on waveforms, that is, Wave-U-Net, by 1.6 points of SDR (signal to distortion ratio). (ii) We propose a new scheme to leverage unlabeled music. We train a first model to extract parts with at least one source silent in unlabeled tracks, for instance without bass. We remix this extract with a bass line taken from the supervised dataset to form a new weakly supervised training example. Combining our architecture and scheme, we show that waveform methods can play in the same ballpark as spectrogram ones.