CITATION — REFERENCE ENTRY

Music Source Separation in the Waveform Domain — arXiv

Revision 419be607-715f-4c67-aafc-ada9155bfdb6 · 11/21/2026, 6:45:56 AM UTC
Key
defossez2019-demucs
Authors
Défossez, Alexandre; Usunier, Nicolas; Bottou, Léon; Bach, Francis
Issued
2019
Type
article
Container
arXiv
Raw CSL JSON
{
  "URL": "https://arxiv.org/abs/1911.13254",
  "note": "Claims verified from v2 (April 28, 2021).",
  "type": "article",
  "title": "Music Source Separation in the Waveform Domain",
  "author": [
    {
      "given": "Alexandre",
      "family": "Défossez"
    },
    {
      "given": "Nicolas",
      "family": "Usunier"
    },
    {
      "given": "Léon",
      "family": "Bottou"
    },
    {
      "given": "Francis",
      "family": "Bach"
    }
  ],
  "issued": {
    "date-parts": [
      [
        2019
      ]
    ]
  },
  "number": "1911.13254",
  "container-title": "arXiv"
}

Claims

  1. Bregman's auditory scene analysis framework spawned a computational counterpart that tries to reproduce the brain's accomplishments, including source separation, by algorithmic means.
    "His framework for auditory scene analysis [Bregman, 1990] spawned its computational counterpart, trying to reproduce or model accomplishments of the brains with algorithmic means [Wang and Brown, 2006], in particular regarding source separation capabilities."
    Locator: page: 1 (sec. 1) · Quote language: en
  2. Demucs, a waveform-to-waveform U-Net with a bidirectional LSTM, reached 6.3 SDR on average on MusDB.
    "We propose instead Demucs, a novel waveform-to-waveform model, with a U-Net structure and bidirectional LSTM. Experiments on the MusDB dataset show that, with proper data augmentation, Demucs beats all existing state-of-the-art architectures, including Conv-Tasnet, with 6.3 SDR on average"
    Locator: page: 1 (Abstract) · Quote language: en
  3. At SiSEC 2018, the best waveform-predicting model averaged an SDR of 3.2 over four sources, against 5.3 for the best spectrogram-masking approach.
    "in the last SiSec Mus evaluation campaign [Stöter et al., 2018], the best model that directly predicts waveforms achieves an average signal-to-noise ratio (SDR) over all four sources of 3.2, against 5.3 for the best approach that predicts spectrograms masks"
    Locator: page: 2 (sec. 1) · Quote language: en
  4. Music source separation isolates the stems (voice, bass, drums and other accompaniment) recorded individually and arranged together to form a song.
    "Source separation for music is the task of isolating contributions, or stems, from different instruments recorded individually and arranged together to form a song. Such components include voice, bass, drums and any other accompaniments."
    Locator: page: 1 (Abstract) · Quote language: en
Available in