README.md

---
task_categories:
- automatic-speech-recognition
language:
- en
- zh
tags:
- multimodal
- avsr
- video-conferencing
- lombard-effect
---

# 🎥 MLD-VC: Multimodal Dataset for Video Conferencing

> **When AVSR Meets Video Conferencing: Dataset, Degradation, and the Hidden Mechanism Behind Performance Collapse (CVPR 2026)**
> 📄 [[Paper\]](https://arxiv.org/abs/2603.22915) | 🤗 [[Hugging Face Dataset\]](https://huggingface.co/datasets/nccm2p2/MLD-VC)

------

## 📌 Overview

**MLD-VC** is the **first multimodal dataset specifically designed for Audio-Visual Speech Recognition (AVSR) in real-world video conferencing (VC) scenarios**.

Unlike traditional AVSR datasets collected in controlled offline environments, MLD-VC explicitly models two critical factors in VC:

- **Transmission Distortions** (compression, speech enhancement, etc.)
- **Human Hyper-expression** (e.g., Lombard effect)

### 🔍 Key Features

- 🎤 **31 speakers**, 22.79 hours of recordings
- 🌐 **4 mainstream VC platforms**
- 🗣️ **Bilingual**: English & Chinese
- 🎧 **Lombard effect simulation** via noise conditions
- 🎥 Multimodal data:
  - Video
  - Audio
  - Facial landmarks
  - text

------

## 🚨 Motivation

Existing AVSR systems show **severe performance degradation in video conferencing**, due to:

- Distribution shift caused by **speech enhancement algorithms**
- Behavioral changes such as **hyper-expression**

MLD-VC is designed to **bridge the gap between offline datasets and real-world VC deployment**.\

------

## 📂 Dataset Structure

The dataset is organized into three aligned modalities:

```
MLD-VC/
├── video/
├── audio/
├── landmarks/
```

Each modality follows the **same hierarchical structure**:

```
<modality>/
└── Online / Offline
    └── speaker_id
        └── platform
            └── sentence_id
                └── clean / 40db / 60db / 80db
```

### 📖 Example

```
video/
└── Online/
    └── speaker_03/
        └── Zoom/
            └── sentence_012/
                ├── clean/
                ├── 40db/
                ├── 60db/
                └── 80db/
```

------

## 🧠 Data Description

### 1. Online vs Offline

- **Offline**:
  - Direct recording (no transmission)
  - Contains hyper-expression (via noise)
- **Online**:
  - Recorded after transmission through VC platforms
  - Includes:
    - Compression
    - Speech enhancement
    - Network effects

------

### 2. Noise Levels (Lombard Effect)

Each sentence is recorded under 4 noise conditions:

| Condition | Description    |
| --------- | -------------- |
| clean     | No noise       |
| 40dB      | Mild noise     |
| 60dB      | Moderate noise |
| 80dB      | Strong noise   |

These simulate **Lombard effect intensity**, inducing hyper-expression.

------

### 3. Platforms

The dataset includes recordings from multiple VC platforms (e.g.):

- Zoom
- Tencent Meeting
- Lark
- DingTalk

------

## ⚠️ Important Notes

### 🔁 Recording Protocol Differences

- In **Offline subset**:
  - **Speakers 2–8**:
    - Recorded on **a single device**, repeated across 4 platforms
  - Other speakers:
    - **DD platform only**, but actually recorded using **4 different devices simultaneously**

👉 This leads to:

- Platform variation ≠ always device variation
- Be careful in **cross-platform generalization experiments**

------

### ❌ Removed Speakers

- **Speaker 0 and 1 have been removed**
  - Due to poor recording quality

------

### 📏 Data Consistency

- All three modalities (`video`, `audio`, `landmarks`) are:
  - **Strictly aligned**
  - Share identical folder structure
  - Can be indexed jointly

------

## 🔬 Recommended Use Cases

MLD-VC is suitable for:

### ✔ AVSR Robustness

- Evaluate performance under real VC conditions

### ✔ Cross-domain Generalization

- Train on Offline → Test on Online

### ✔ Multimodal Learning

- Audio-visual fusion
- Landmark-based modeling

### ✔ Distribution Shift Analysis

- Study impact of:
  - Speech enhancement
  - Lombard effect

------

## 📊 Key Findings (from the paper)

- AVSR models suffer **massive degradation in VC**
- **Speech enhancement** is the main cause of audio distribution shift
- **Lombard effect ≈ VC distortion (in feature space)**
- Landmark-based features are **more stable than image features**
- Fine-tuning on MLD-VC reduces CER by **17.5%**

------

## 📎 Citation

If you find this dataset useful, please cite:

```bibtex
@inproceedings{huang2026mldvc,
  title={When AVSR Meets Video Conferencing: Dataset, Degradation, and the Hidden Mechanism Behind Performance Collapse},
  author={Huang, Yihuan and Xue, Jun and Liu, Jiajun and Li, Daixian and Zhang, Tong and Yi, Zhuolin and Ren, Yanzhen and Li, Kai},
  booktitle={CVPR},
  year={2026}
}
```

------

## 🙏 Acknowledgements

This work is supported by:

- National Natural Science Foundation of China
- DiDi Chuxing Group

------

## 📬 Contact

If you have questions, feel free to contact:

- **Yihuan Huang**: [yihuanhuang@whu.edu.cn](mailto:yihuanhuang@whu.edu.cn)

------

## ⭐ Star This Repo

If you find MLD-VC helpful, please consider giving a ⭐!