README.md

---
task_categories:
- audio-to-audio
tags:
- music
pretty_name: MulTTiPop
size_categories:
- n<1K
license: cc-by-4.0
configs:
- config_name: default
  data_files:
  - split: dev
    path: dev.json
  - split: test
    path: test.json
viewer: true
---

<img src="./multtipoplogo.png" style="float:left" alt="MulTTiPop Logo" width="60"/>

# MulTTiPop: A Multitrack Transcription Dataset for Pop Music

[Nathan Pruyne](https://nathanpruyne.com), 
[Benjamin Stoler](https://benstoler.com/), 
[William Chen](https://wanchichen.github.io), 
[Chien-yu Huang](https://chien-yu.com/), 
[Shinji Watanabe](https://sites.google.com/view/shinjiwatanabe), 
[Chris Donahue](https://chrisdonahue.com/)

🎼**Preview the Data**: [https://gclef-cmu.org/multtipop](https://gclef-cmu.org/multtipop)

📄**Read the Paper**: [https://arxiv.org/pdf/2607.08756](https://arxiv.org/pdf/2607.08756)

**MulTTiPop** is a dataset of segments of pop music, sourced from TheoryTab, and aligned multitrack MIDI transcriptions, sourced from the Lakh MIDI Dataset. MulTTiPop contains 572 total segments and approximately 3.5 hours of total data.

We release MulTTiPop as pairs of aligned multitrack MIDI files and associated metadata files with information about the audio. Specifically, the metadata contains YouTube video IDs and timestamps within the video for the audio that corresponds to the MIDI labels.

We make 2 key recommendations for the use of MulTTiPop:

1. Researchers should only obtain the relevant segments of the original audio.
2. Researchers should only use MulTTiPop for model evaluation, not training.

## Splits

MulTTiPop is split in an approximately 3:7 ratio between `dev` and `test` splits. The `dev` split is intended for development, and the `test` split for model evaluation. As MulTTiPop is not intended for model training, we do not provide a train or validation split. 

## Uses

MulTTiPop is intended for use for the evaluation of multitrack automatic music transcription (AMT) systems.

### Direct Use

For the evaluation of AMT systems such as [MT3](https://github.com/magenta/mt3) and [YourMT3+](https://huggingface.co/spaces/mimbres/YourMT3)

### Out-of-Scope Use

MulTTiPop is not intended for model training due to both its size and lack of audio licensing.

## Dataset Structure

Each folder in the `dev` and `test` splits contain two files: an aligned MIDI file (`aligned.mid`) and a metadata file (`meta.json`).

### Metadata Fields

- `"id"`: Internal dataset ID
- `"artist"`, `"name"`, `"section"`, `"tempo"`, `"meter"`, `"key"`: Audio metadata sourced from TheoryTab
- `"youtube"`: YouTube video associated with the song. Contains the following subfields:
  - `"id"`: YouTube video ID (video can be accessed at `https://youtube.com/watch?v=VIDEO_ID`)
  - `"start"`: Start time of the video segment, in seconds
  - `"end"`: End time of the video segment, in seconds
- `"chosen_method"`: Starting anchor beat selection method for audio-MIDI alignment. See our paper for more details.
- `"chosen_beat_idx"`: Starting beat number in the LMD MIDI file
- `"chosen_beat_time"`: Starting time in the LMD MIDI file in seconds
- `"chosen_lmd_midi_filename"`: Filename in the LMD MIDI dataset
- `"chosen_candidate"`: Additional information used for audio-MIDI alignment
- `"lmd_beats"`: Timestamps of beats within the LMD MIDI, used for audio-MIDI alignment
- `"audio_beats"`: Timestamps of beats within the audio, used for audio-MIDI alignment
- `"aligned_midi_checksum"`: SHA-256 checksum of the associated `aligned.mid` file
- `"lmd_midi_checksum"`: SHA-256 checksum of the original LMD MIDI file
- `"split_name"`: `dev` or `test`
- `"year"`: Publication year of the song, according to Claude Sonnet
- `"audio_length"`: Total length of the associated audio

### Source Data

MulTTiPop segments are sourced from [TheoryTab](https://hooktheory.com/theorytab).

Multitrack MIDI is adapted from the [Lakh MIDI Dataset](https://colinraffel.com/projects/lmd/).


## Bias, Risks, and Limitations

MulTTiPop is an evaluation dataset for automatic music transcription models on multitrack pop music. 
It is intended to provide a representative sample of songs that we hypothesize users of AMT systems would want to transcribe. 
MulTTiPop is not intended for the training of AMT systems or other machine learning models, especially generative models, both because of the small size of the dataset and due to it being comprised of copyrighted songs.

While MulTTiPop is intended to give a representative sample of music that users may wish to transcribe, there is a strong Western bias in our data preparation and presentation. 
The transcriptions we provide are based on Western harmony and notation. 
By sourcing our annotations of interest from HookTheory, an American company with a primarily Western audience, the songs included in MulTTiPop overwhelmingly represent the transcription interests of a Western, not global audience. 
This is reflected in the contents of the data too, as the artists included are predominantly American and are popular in the American pop music landscape. 
Because of these limitations, we recommend that model results on MulTTiPop should be considered largely a reflection of their performance on music in the American Western popular music canon, and not representative of performance on music as a whole.

## Citation

If you use MulTTiPop in your research, please cite our ArXiv paper:

```
@article{pruyne2026multtipop,
  title={MulTTiPop: A Multitrack Transcription Dataset for Pop Music},
  author={Pruyne, Nathan and Stoler, Benjamin and Chen, William and Huang, Chien-yu, and Watanabe, Shinji and Donahue, Chris},
  journal={arXiv preprint arXiv:2607.08756},
  year={2026}
}
```

## License and Attribution

MulTTiPop is distributed with a [CC-BY 4.0 License](https://creativecommons.org/licenses/by/4.0/).

The aligned MIDI files provided in MulTTiPop are adapted from the [Lakh MIDI Dataset](https://colinraffel.com/projects/lmd/) by Colin Raffel, which is distributed with a [CC-BY 4.0 License](https://creativecommons.org/licenses/by/4.0/).