CITATION — REFERENCE ENTRY
Hybrid Spectrogram and Waveform Source Separation — Proceedings of the MDX Workshop
- Key
- defossez2021-hybrid-demucs
- Authors
- Défossez, Alexandre
- Issued
- 2021
- Type
- paper-conference
- Container
- Proceedings of the MDX Workshop
Raw CSL JSON
{
"URL": "https://arxiv.org/abs/2111.03600",
"note": "Claims verified from arXiv 2111.03600 v3.",
"type": "paper-conference",
"title": "Hybrid Spectrogram and Waveform Source Separation",
"author": [
{
"given": "Alexandre",
"family": "Défossez"
}
],
"issued": {
"date-parts": [
[
2021
]
]
},
"container-title": "Proceedings of the MDX Workshop"
}
Claims
-
Different representations lead to different artifacts: spectrogram methods can make drum and bass attacks sound hollow, while vocals separated by waveform Demucs suffer from crunchy static noise.
"Different representations will lead to different artifacts, some being more noticeable for the drums and bass (phase inconsitency for spectrogram methods will make the attack sounds hollow), while others are more noticeable for the vocals (vocals separated by Demucs suffer from crunchy static noise)"
-
Music source separation research has recently focused on supervised separation of four instruments: drums, bass, vocals and other accompaniment.
"Work on music source separation has recently focused on the task of separating 4 well defined instruments in a supervised manner: drums, bass, vocals and other accompaniments."
-
Hybrid Demucs ranked first at the MDX competition among systems trained only on MusDB, with 7.32 dB SDR.
"Hybrid Demucs ranked 1st at the MDX competition when trained only on MusDB, with 7.32 dB of SDR, and 2nd with extra training data allowed."
-
Hybrid Demucs performs end-to-end hybrid source separation, letting the model decide which domain suits each source and even combining both.
"In this work, we show how to perform end-to-end hybrid source separation, letting the model decide which domain is best suited for each source, and even combining both."
Available in