Chamber Ensemble Generator and CocoChorales Dataset

Blog | Paper | Dataset Page | Sample Website | Paper With Code | ICML Workshop Poster

Chamber Ensemble Generator is a music dataset generation pipeline consist of Coconet and MIDI-DDSP. Chamber Ensemble Generator can generates 4-part Bach Chorales performance audio with aligned notes, note expressions, synthesis parameters (f0, amplitudes, etc) and stems while enabling rich variations in generative process.

We use Chamber Ensemble Generator to generate a large-scale dataset CocoChorlaes. CocoChorales consists of 240,000 pieces in a total duration of 1400 hours, with aligned notes, note expressions, synthesis parameters (f0, amplitudes, etc) and stems. For details of CocoChorales, please check our paper.

CocoChorales Samples

For samples and some statistics of CocoChorales dataset, please check our sample website.

Dataset Creation

Please check data_pipeline.md for detials of Chamber Ensemble Geneartor and how to create CocoChorales Dataset or your won dataset.

CocoChorales Format

Please check data_format.md for the format of the CocoChorlaes dataset.

Dataset Download

Simply run data_download/download_cocochorales_full.sh to download the full CocoChorales dataset consists of 240k samples.

The script will download all the data types in the CocoChorales dataset to the cocochorales_full_v1_zipped folder it creates under current directory. Please see data_format.md for details of each type. You could comment the lines of the data download you don't want to download.

Below are the size of 5 types of data in CocoChorales full dataset:

• main_dataset: 569G.
• note_expression: 1.3G.
• synthesis_parameters: 2.3T.
• f0: 6.1G.
• metadata: 279M.
• total: 2.9T.

You could run data_download/download_cocochorales_tiny.sh to download a tiny version of CocoChorales dataset. The script will download all the data types in the CocoChorales tiny dataset to the cocochorales_tiny_v1_zipped folder it creates under current directory The tiny version contains 24k training samples, 8k validation samples and 8k test samples.

Here is md5 of the CocoChorales dataset.

Dataset Extraction

You could run data_download/extract_tars.py to extract the downloaded tar files.

The code uses tar command in command line to extract the tar files and uses pbzip2 compressor. If you do not have pbzip2, here is how to install that on macOS or Linux. If you are using Windows, you might consider use other software to extract the tar files and put the extracted files to the correct directory.

python data_download/extract_tars.py --data_dir <dir_to_cocochorales_full_v1_zipped> --output_dir <dir_to_cocochorales_full>

Citation

If you use CocoChorales dataset in your research, please consider cite our paper:

@article{wu2022chamber,
  title = {The Chamber Ensemble Generator: Limitless High-Quality MIR Data via Generative Modeling},
  author = {Wu, Yusong and Gardner, Josh and Manilow, Ethan and Simon, Ian and Hawthorne, Curtis and Engel, Jesse},
  journal={arXiv preprint arXiv:2209.14458},
  year = {2022},
}