---
license: apache-2.0
---

# LLM-Generated Hungarian Conversations

This dataset contains structured Hungarian conversations generated with multiple large language model families for the study **“Efficient ASR Training with Conversations that Never Happened.”** Paper: [arXiv link](https://arxiv.org/abs/2606.03957v1)

Each model is provided as a separate Parquet file. The dataset contains the generated textual conversations and associated scenario and participant metadata; it does not contain synthesized audio.

## Dataset structure

Each row corresponds to one sentence within a conversational turn.

| Field | Description |
|---|---|
| `conv_id` | Unique conversation identifier |
| `turn_id` | Turn index within the conversation |
| `speaker` | Speaker identifier (`speaker_1` or `speaker_2`) |
| `source_file` | Original generated text-file name |
| `topic_id` | Topic identifier |
| `scenario_id` | Scenario index |
| `topic` | Conversation topic |
| `participant` | Participant metadata, including age, gender, and occupation |
| `interviewer` | Interviewer metadata, including age, gender, and role |
| `sentence_id` | Sentence index within the turn |
| `text` | Hungarian sentence text |

Rows sharing the same `conv_id` belong to the same conversation. Sentences should be ordered by `turn_id` and then by `sentence_id`.

## Citation

When using this dataset, please cite the associated paper:

```bibtex
@misc{gedeon2026efficientasrtrainingconversations,
      title={Efficient ASR Training with Conversations that Never Happened}, 
      author={Máté Gedeon and Péter Mihajlik},
      year={2026},
      eprint={2606.03957},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.03957}, 
}
```