---
license: cc-by-4.0
language: en
task_categories:
  - audio-classification
tags:
  - end-turn-detection
  - turn-taking
  - spoken-dialogue
  - conversation
pretty_name: OpenETD Metadata
size_categories:
  - 100K<n<1M
---

# OpenETD metadata

Train / dev / test split CSVs for the OpenETD dataset released with the ACL 2026 Findings paper *Speculative End-Turn Detector for Efficient Speech Chatbot Assistant* ([arXiv:2503.23439](https://arxiv.org/abs/2503.23439)).

Code: https://github.com/HJ-Ok/OpenETD

## Split sizes

| Split | Real files | Real hours | Synthetic files | Synthetic hours |
|-------|------------|------------|-----------------|-----------------|
| train | 6,290      | 117.2      | 96,773          | 116.8           |
| dev   |   899      |  16.2      | 12,840          |  15.8           |
| test  | 1,798      |  32.4      | 12,868          |  15.7           |

## Columns

| Column           | Description                                                            |
|------------------|------------------------------------------------------------------------|
| `file_path`      | Relative path to the audio file (resolve locally).                     |
| `pause_times`    | Interval list `(start, end), ...` of within-speaker pauses (seconds).  |
| `gap_times`      | Interval list `(start, end), ...` of between-speaker gaps (seconds).   |
| `contains_pause` | Boolean, whether the file contains any pause.                          |
| `contains_gap`   | Boolean, whether the file contains any gap.                            |
| `label`          | Type of the final silence (`Pause` or `Gap`); used for the binary task.|
| `platform`       | (Real only) `buckeye` or `youtube`.                                    |
| `kfold`          | (Synthetic only) k-fold assignment used for pause/gap label generation.|

## Audio files

**Audio is NOT included** in this repository — we redistribute only the annotations and split assignments. To obtain the audio:

- **Buckeye audio**: obtain from the [Buckeye Corpus](https://buckeyecorpus.osu.edu/) maintainers under their Academic License, then place files under `data/real/audio/buckeye_full/`.
- **YouTube audio**: download with the helper script in [OpenETD repository](https://github.com/HJ-Ok/OpenETD) (`scripts/prepare_data.sh`).
- **Synthetic audio**: regenerate on your own Google Cloud account using `data/synthetic_pipeline/generate.py` in the OpenETD repository.

## Quick start

```python
from datasets import load_dataset
ds = load_dataset("HJOK/openetd-metadata", data_files={
    "real_train":  "real/train.csv",
    "real_valid":  "real/valid.csv",
    "real_test":   "real/test.csv",
    "syn_train":   "synthetic/train.csv",
    "syn_valid":   "synthetic/valid.csv",
    "syn_test":    "synthetic/test.csv",
})
print(ds["real_test"][0])
```

## License

- Annotations (this repository): **CC BY 4.0**
- Code in the OpenETD GitHub repository: **MIT**
- External audio sources retain their original licenses (see `DATA_LICENSES.md` in the GitHub repo).

## Citation

```bibtex
@inproceedings{ok2026speculativeetd,
  title     = {Speculative End-Turn Detector for Efficient Speech Chatbot Assistant},
  author    = {Ok, Hyunjong and Yoo, Suho and Lee, Jaeho},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year      = {2026},
  url       = {https://arxiv.org/abs/2503.23439}
}
```