schism-audio/dnr-v2

Divide and Remaster v2

This repository mirrors version 2.0 of the Divide and Remaster (DnR) dataset from Zenodo.

DnR is a source separation dataset for separating a monaural soundtrack mixture into speech, music, and sound effects/background stems. It is composed of artificial mixtures using LibriSpeech, Free Music Archive, and FSD50K source audio.

The upstream release contains 4866 one-minute mixtures split into tr (3406), cv (487), and tt (973). Each mixture directory contains mix.wav, music.wav, speech.wav, sfx.wav, and annots.csv. Audio is 32-bit WAV at 44.1 kHz. The compressed release is split into eleven archive chunks totaling about 116.4 GB, and the extracted dataset is about 200 GB.

This mirror preserves the upstream split archive files and exposes metadata.csv for the dataset viewer.

Source

• Zenodo: https://zenodo.org/records/6949108
• DOI: 10.5281/zenodo.6949108
• License: Creative Commons Attribution 4.0 International
• Project page: https://cocktail-fork.github.io/
• Generation code: https://github.com/darius522/dnr-utils

Reconstructing the Dataset

Download all files under archives/ into the same directory, then run:

cat dnr_v2.tar.gz.* > dnr_v2.tar.gz
tar -xf dnr_v2.tar.gz

Citation

@INPROCEEDINGS{petermann2021cfp,
  author={Petermann, Darius and Wichern, Gordon and Wang, Zhong-Qiu and Roux, Jonathan Le},
  booktitle={ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  title={The Cocktail Fork Problem: Three-Stem Audio Separation for Real-World Soundtracks},
  year={2022},
  pages={526-530},
  doi={10.1109/ICASSP43922.2022.9746005}
}