# CocoChorales: 240,000 Synthesized Four-Part Chorale Mixtures with Per-Instrument Stems, MIDI and Note Expression (Google Magenta)

## Facts

- Provider: Google Magenta (Yusong Wu, Josh Gardner, Ethan Manilow, Ian Simon, Curtis Hawthorne, Jesse Engel)
- License: CC-BY-4.0 (https://magenta.tensorflow.org/datasets/cocochorales)
- Stored at: Google Cloud Storage, gs://magentadata/datasets/cocochorales (us)
- Access: public, readable without an account
- Category: Speech & Audio
- Source: https://arxiv.org/abs/2209.14458
- Page: https://scigantic.com/archive/e6515d6c-eefd-586d-8c0d-9b5abf2ecdaf
- Keywords: Speech & Audio, Google Cloud Storage

## About

CocoChorales is a music dataset of over 1,400 hours of audio mixtures of four-part chorales performed by 13 instruments, all synthesized with generative models. A trained Coconet model first generated four-part chorales in the style of J.S. Bach (stored as MIDI, with tempo and random note timing variations), then MIDI-DDSP, trained on the URMP dataset, rendered each part as audio. Ensembles are String (violin, violin, viola, cello), Brass (trumpet, French horn, trombone, tuba), Woodwind (flute, oboe, clarinet, bassoon) and Random (each soprano, alto, tenor and bass part drawn from a list of instruments). The dataset has 240,000 examples, 60,000 mixes per ensemble type, each with its own train/validation/test split; audio is 16 kHz 16-bit PCM. Every example contains the mixture, the source audio of all four instruments, per-source gain, MIDI, per-note expression annotations (volume, vibrato, brightness and others) and synthesis parameters (fundamental frequency f0, harmonic amplitudes, filtered noise) with the pitch augmentation applied. Because the data come from generative models, annotations match the audio exactly. The folder cocochorales_full_v1_zipped/ splits the zipped data into main_dataset/, metadata/, note_expression/, synthesis_parameters/ and f0/, with an MD5 list; about 2.9 TB compressed. Used for music source separation, automatic music transcription and multi-f0 estimation. License: CC BY 4.0. Cite Wu et al., The Chamber Ensemble Generator: Limitless High-Quality MIR Data via Generative Modeling, arXiv:2209.14458.

## Load the data

```python
import gcsfs
fs = gcsfs.GCSFileSystem(token='anon')
fs.ls('magentadata/datasets/cocochorales/', detail=False)[:10]
```

Lists the files with anonymous access. Runs on any machine, with no cloud account.

## Data structure

Example file: `datasets/cocochorales/cocochorales_full_v1_zipped/f0/test/1.tar.bz2`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/e6515d6c-eefd-586d-8c0d-9b5abf2ecdaf
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/e6515d6c-eefd-586d-8c0d-9b5abf2ecdaf
- Over MCP: `get_schema_card` with id `e6515d6c-eefd-586d-8c0d-9b5abf2ecdaf` at https://api.scigantic.com/mcp

## Related topics

- [Speech & Audio datasets](https://scigantic.com/topics/speech-and-audio)
