Divide and Remaster (DnR)

Introduction:

Divide and Remaster (DnR) is a source separation dataset for training and testing algorithms that separate a monaural audio signal into speech, music, and sound effects/background stems. The dataset is composed of artificial mixtures using audio from the librispeech, free music archive (FMA), and Freesound Dataset 50k (FSD50k). We introduce it as part of the Cocktail Fork Problem paper.

At a Glance:

• The size of the unzipped dataset is ~200GB
• Each mixture is 60 seconds long and sources are not fully overlapped
• Audio is encoded as 32-bit .wav files at a sampling rate of 44.1 kHz
• The data is split into training `tr` (3406 mixtues), validation `cv` (487 mixtures) and testing `tt` (973 mixtures) subsets
• The directory for each mixture contains four .wav files, mix.wav, music.wav, speech.wav, sfx.wav, and annots.csv which contains the metadata for the original audio used to compose the mixture (transcriptions for speech, sound classes for sfx, and genre labels for music)

Other Resources:

Demo examples and additional information are available at: https://cocktail-fork.github.io/

For more details about the data generation process, the code used to generate our dataset can be found at the following: https://github.com/darius522/dnr-utils

Fix Brought to V2:

V1 contained some errors as for the speech annotations, which in some cases did not match with their associated audio utterances. This was caused by some of the speech utterances being truncated to fit the length of their associated DnR mixtures, while their transcriptions were reported in their entirety. To address the issue, we discarded the possibility of any utterances being truncated during the DnR creation process. Since the number of test-set mixtures is determined such that we exhaust all utterances from the LibriSpeech TEST-CLEAN set twice, the test-set grew larger in V2 (from 652 to 973 mixtures). To maintain the same split proportions, the training and validation sets have been increased accordingly as well (3406 and 487, respectively).

All the results in the camera-ready paper have also been updated to reflect these changes and using the DnR V2.

In this version, we split the datasets into smaller chunks to ease-up the download process.

Contact and Support:

Have an issue, concern, or question about DnR ? If so, please open an issue here.

For any other inquiries, feel free to shoot an email at: firstname.lastname@gmail.com, my name is Darius Petermann ;)

Download:

DnR V2 is split into smaller ~10GB chunks to ease-up the download process. First download all the chunks into a single directory:

dnr_v2.tar.gz.00
dnr_v2.tar.gz.01
dnr_v2.tar.gz.02
dnr_v2.tar.gz.03
dnr_v2.tar.gz.04
dnr_v2.tar.gz.05
dnr_v2.tar.gz.06
dnr_v2.tar.gz.07
dnr_v2.tar.gz.08
dnr_v2.tar.gz.09
dnr_v2.tar.gz.10

From the same directory run the following to compile all the chunks into a single .tar file:

cat dnr_v2.tar.gz.* >dnr_v2.tar.gz

Finally untar the resulting file:

tar -xf dnr_v2.tar.gz

Citation:

If you use DnR please cite our paper in which we introduce the dataset as part of the Cocktail Fork Problem:

@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},
  volume={},
  number={},
  pages={526-530},
  doi={10.1109/ICASSP43922.2022.9746005}
}