BeTraC/betrac-2026

---
dataset_info:
- config_name: default
features:
- name: opus
dtype: binary
- name: transcript.txt
dtype: string
- name: soap.txt
dtype: string
- name: json
dtype: string
splits:
- name: validation
num_examples: 400
num_bytes: 469422080
- name: train
num_examples: 7200
num_bytes: 8661381120
license: cc-by-4.0
task_categories:
- automatic-speech-recognition
- summarization
language:
- en
tags:
- medical
- doctor-patient
- webdataset
- soap-notes
- betrac
size_categories:
- 1K<n<10K
pretty_name: BeTraC 2026 - DoPaCo Audio Dataset
---

# BeTraC 2026 - Synth-DoPaCo Audio Dataset

Synthetic doctor-patient conversations with audio, transcripts, dialog metadata, and SOAP note summaries.

## Dataset Splits

| Split | Dialogs | Shards | Size |
|---|---|---|---|
| `dev` | 400 | 1 | 469 MB |
| `train` | 7,200 | 9 | 8.7 GB |

## File Format

This dataset uses the [WebDataset](https://github.com/webdataset/webdataset) format (tar archives).
Each sample contains 4 files sharing the same key (e.g., `dialog_0060_0120`):

| Extension | Content |
|---|---|
| `.opus` | Opus-compressed audio (16 kHz mono) |
| `.transcript.txt` | Full transcript of the doctor-patient dialog |
| `.json` | Dialog metadata (personas, generation parameters, dialog turns) |
| `.soap.txt` | Target SOAP note summary |

## Usage

```python
import webdataset as wds
import json

dataset = wds.WebDataset("path/to/dev-00000.tar", shardshuffle=False)

for sample in dataset:
key = sample["__key__"]
audio_bytes = sample["opus"] # raw Opus bytes
transcript = sample["transcript.txt"].decode("utf-8")
soap_note = sample["soap.txt"].decode("utf-8")
metadata = json.loads(sample["json"])
print(f"{key}: {len(audio_bytes)} bytes audio, {len(transcript)} chars transcript")
break
```

### Streaming from Hugging Face Hub

```python
import webdataset as wds
from huggingface_hub import get_token

token = get_token()
url = "https://huggingface.co/datasets/BeTraC/betrac-2026/resolve/main/data/train-{00000..00008}.tar"
url = f"pipe:curl -s -L {url} -H 'Authorization:Bearer {token}'"

dataset = wds.WebDataset(url, shardshuffle=False)
for sample in dataset:
print(sample["__key__"])
break
```

### Decoding Audio

The `.opus` files are Ogg/Opus containers. Decode with `soundfile`, `torchaudio`, or `ffmpeg`:

```python
import soundfile as sf
import io

audio_data, sample_rate = sf.read(io.BytesIO(sample["opus"]))
```

## License

This dataset is released under [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/).

## Citation

Please the cite forthcoming paper.

## Acknowledgments

The Synth-DoPaCo dataset used in BeTraC 2026 was created by the Play-Your-Part team during the [JSALT 2025](https://jsalt2025.fit.vut.cz/) workshop, organized by the Center for Language and Speech Processing at Johns Hopkins University and held at Brno University of Technology.