---
language:
- zh
license: apache-2.0
task_categories:
- automatic-speech-recognition
viewer: true
dataset_info:
- config_name: User
  features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 16000
  - name: text
    dtype: string
  - name: duration
    dtype: float64
  - name: chat_id
    dtype: string
  splits:
  - name: train
    num_bytes: 889743413
    num_examples: 1008
  - name: eval
    num_bytes: 235780103
    num_examples: 253
- config_name: Agent
  features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 16000
  - name: text
    dtype: string
  - name: duration
    dtype: float64
  - name: chat_id
    dtype: string
  splits:
  - name: train
    num_bytes: 3749359214
    num_examples: 4552
  - name: eval
    num_bytes: 930931836
    num_examples: 1138
configs:
- config_name: User
  data_files:
  - split: train
    path: User/train/*
  - split: eval
    path: User/eval/*
- config_name: Agent
  data_files:
  - split: train
    path: Agent/train/*
  - split: eval
    path: Agent/eval/*
tags:
- medical
- healthcare
size_categories:
- 1K<n<10K
---

## MMedFD: A Real-World Healthcare Benchmark for Multi-Turn Full-Duplex Automatic Speech Recognition

### 📄 **Preprint**: [MMedFD](https://arxiv.org/abs/2509.19817) — For the complete benchmark construction pipeline, evaluation methodology, dataset specifications, and additional implementation details, please refer to the preprint.


### ⚠️Data Availability
Full access requires internal approval and a research-only data use agreement. 
🚫 Non-Commercial Use
This dataset is provided **for non-commercial research and education only**. **Commercial use is prohibited.**  
Researchers who wish to request full access may contact yangxiao.wxy@antgroup.com with a brief description of their affiliation, project goals, intended use, and data protection plan. Only de-identified data may be shared, and redistribution is prohibited.

## 🗂️ Data Release & Access

- **Public release (partial subset)**: We release **only a portion of the data used for this benchmark’s training and evaluation**. This Lite subset **differs in amount and coverage** from our internal full dataset and is **not** a drop-in replacement for the complete data.
- **What’s included**: A **reduced selection** of dialogues/audio/text sufficient to reproduce the reported benchmark protocol at a smaller scale.
- **Not included**: Additional sessions, higher-fidelity artifacts, and full validation/test coverage remain internal.

## 🔒 Privacy, Safety & Redaction

- **Privacy-preserving audio**: To protect speaker privacy, all audio has been **re-synthesized via TTS** (privacy-preserving re-encoding). This process **obfuscates speaker identity and acoustic biomarkers** while preserving task-relevant linguistic content for modeling.
- **Real-world dialogs**: The **dialogue content originates from real-world collections**. However, **sensitive spans** (e.g., direct identifiers, highly specific personal details) are **automatically redacted by an LLM-based filter** before release.
- **Residual risk**: Despite these protections, **re-identification attempts are prohibited**. Please do not try to recover original identities or link samples to outside sources.


## 📑 How to Cite

If this code or our benchmark is useful for your research, please consider citing our paper:
```bibtex
@misc{chen2025mmedfdrealworldhealthcarebenchmark,
      title={MMedFD: A Real-world Healthcare Benchmark for Multi-turn Full-Duplex Automatic Speech Recognition}, 
      author={Hongzhao Chen and XiaoYang Wang and Jing Lan and Hexiao Ding and Yufeng Jiang and MingHui Yang and DanHui Xu and Jun Luo and Nga-Chun Ng and Gerald W. Y. Cheng and Yunlin Mao and Jung Sun Yoo},
      year={2025},
      eprint={2509.19817},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2509.19817}, 
}
```