HEALTHDIAL

A Multilingual and Multi-Parallel Spoken Dialogue Dataset for Knowledge-Grounded Information Seeking

HEALTHDIAL is a large-scale, multilingual, multi-parallel spoken dialogue dataset for developing and evaluating retrieval-augmented generation (RAG)-based spoken dialogue systems.

• 6,000 information-seeking dialogues (1,500 per language) across Arabic, Chinese, English, and Spanish.
• 163 hours of user speech recorded by native speakers of diverse dialects, plus 208 hours of system speech.
• 12,045 unique WHO knowledge snippets (6,472 fully parallel across all four languages) grounding the dialogues.
• Speaker-level demographic and sociolinguistic annotations (gender, age, region of origin, primary language, education level).

The dataset is hosted on HuggingFace at cambridgeltl/HealthDial. This repository hosts the code released alongside the paper.

---

What is in this repository

Directory	Purpose
benchmark/	Baseline code for the four benchmark tasks: ASR, TTS, retrieval-turn classification, knowledge retrieval (text-to-text and speech-to-text), and knowledge filtering. Includes the HealthDialogueDatabase loader and a HuggingFace download helper.
annotation_tool/	Web platform used to collect the human-authored user utterances (Figure 8 in the paper). React client + Flask API + audio recording / Whisper transcription pipeline.
human_eval_tool/	TAM2-based human-evaluation platform (Figure 12). React participant interface + Flask backend + MongoDB + a dummy Socket.IO dialogue server for local testing.

Each subdirectory has its own README with installation and usage instructions.

---

Quick start

To reproduce the benchmark numbers in the paper:

git clone https://github.com/cambridgeltl/healthdial.git
cd healthdial/benchmark

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# Authenticate once for HuggingFace if needed:
huggingface-cli login

# Pull the dataset (audio + dialogue JSONs):
python download_hf.py --output-dir data

# Run a benchmark, e.g. BM25 on English:
cd retrieval/evaluation
python run_retrieval.py --language eng --retriever bm25 --sample no

See benchmark/README.md for the full list of supported benchmarks and per-task instructions.

To collect your own dialogues with the same annotation interface, see annotation_tool/README.md. To run a TAM2-style user study against your own dialogue system, see human_eval_tool/README.md.

---

Citation

If you use HEALTHDIAL in your work, please cite:

@inproceedings{hu2025healthdial,
  title     = {{Dial HealthDial for Advice}: A Multilingual and Multi-Parallel Spoken Dialogue Dataset for Knowledge-Grounded Information Seeking},
  author    = {Hu, Songbo and Liu, Yinhong and Zhou, Ej and Razumovskaia, Evgeniia and Wang, Xiaobin and Fraser, Alexander and Vuli{\'c}, Ivan and Korhonen, Anna},
  year      = {2025},
}

---

License & ethical use

• Code in this repository (all three sub-directories) is released under the MIT License.
• Audio data on HuggingFace is released under a separate, non-commercial data use agreement that prohibits voice cloning and re-identification of annotators. See the dataset page for full terms.
• HEALTHDIAL is released as a language resource for research, not a clinical tool. The knowledge snippets come from the WHO website but have not been validated by healthcare professionals. Do not deploy this in a clinical setting without expert review.

---

Acknowledgements

This work is supported by the Cambridge–LMU Strategic Partnership grant, the UKRI Frontier Research Grant EP/Y031350/1 EQUATE (Anna Korhonen), the Cambridge International Scholarship (Songbo Hu), and a Royal Society University Research Fellowship (Ivan Vulić, no. 221137).