# Beethoven Sonatas Dataset

Beethoven is a raw audio waveform dataset used in the paper "It's Raw! Audio Generation with State-Space Models". It has been used primarily as a source of single instrument piano music for training music generation models at a small scale.

The dataset was originally introduced in the SampleRNN paper by Mehri et al. (2017) and download details from the original paper can be found at https://github.com/soroushmehr/sampleRNN_ICLR2017/tree/master/datasets/music. Here, we provide a more convenient download of a processed version of the dataset in order to standardize future use.


We include two versions of the dataset:
- `beethoven.zip` is a zip file containing 4328 8-second audio clips sampled at 16kHz. These were generated by first joining all the piano sonatas, and then splitting the track into 8-second chunks. This data can also be used with the https://github.com/HazyResearch/state-spaces repository to reproduce SaShiMi results, and was the dataset used in the paper.
- `beethoven_raw.zip` contains the raw audio tracks, sampled at 16kHz.

We recommend (and follow) the following train-validation-test split for the audio files in `beethoven.zip` (we attempted to recreate the splits from the SampleRNN work as closely as possible):

- `0.wav` to `3807.wav` for training
- `3808.wav` to `4067.wav` for validation
- `4068.wav` to `4327.wav` for testing

You can use the following BibTeX entries to appropriately cite prior work if you decide to use this in your research:
```
@article{goel2022sashimi,
  title={It's Raw! Audio Generation with State-Space Models},
  author={Goel, Karan and Gu, Albert and Donahue, Chris and R\'{e}, Christopher},
  journal={arXiv preprint arXiv:2202.09729},
  year={2022}
}

@inproceedings{mehri2017samplernn,
  title={SampleRNN: An Unconditional End-to-End Neural Audio Generation Model},
  author={Mehri, Soroush and Kumar, Kundan and Gulrajani, Ishaan and Kumar, Rithesh and Jain, Shubham and Sotelo, Jose and Courville, Aaron and Bengio, Yoshua},
  booktitle={International Conference on Learning Representations},
  year={2017}
}
```