medataset

Dataset Description

Synthetic conversational dataset between doctors and patients in Vietnamese, featuring realistic code-switching with English medical terms.

• Languages: vi, en (code-switched)
• License: cc-by-nc-sa-4.0
• Synthetic Source: Generated using LLMs with plan-first SOAP outlines and validated via TTS.

Quality Metrics & Statistics

• Total Conversations: 2423
• Average UTMOS: 4.17
• Average Roundtrip WER: 0.07%
• Code-Switch Ratio: ~0.0%

(Full detailed evaluation metrics can be found in the reports/ folder of this repository).

PII/Safety Statement

This dataset contains purely synthetic clinical dialogues and does not contain any real Patient Identifying Information (PII). It is designed for research and educational purposes and should not be used for clinical decision-making.

Generation Method

The dialogues were generated using a two-step LLM pipeline:

1. Generation of SOAP outlines based on ICD-10 TT06 chapters.
2. Dialogue expansion with code-switching constrained by a curated allow-list of English medical terms.
3. Audio synthesized using varied regional TTS models.

Limitations

As a synthetic dataset, it may not perfectly represent the full range of natural clinical conversations or dialectal variations, and could contain model-generated biases.

Dataset Structure

The dataset is structured as a standard Hugging Face Audio dataset using Parquet format. Each row contains:

• file_name (Audio feature): The waveform data and sampling rate.
• transcript: The full text of the conversation.
• meta: Detailed metadata including quality metrics (UTMOS, WER), ICD-10 chapters, and code-switching statistics.

References

Based on the ICD-10 TT06 diagnostic classification.