metadata
license: cc-by-nc-sa-4.0
EmoDialogCN: A Multimodal Mandarin Dyadic Dialogue Dataset of Emotions
BackGround and Dataset Statistics
Official repository of the EmoDialogCN dataset – the first large-scale, high-quality Mandarin auditory-visual-emotion dialogue dataset featuring 21,800 spontaneous dialogues (over 400 hours) with synchronized video and audio.
- Actor diversity: 119 professional actors spanning varied demographics (age, gender, etc.).
- Emotion coverage: 18 discrete emotional categories with detailed annotations.
- Scenario richness: 20 everyday real-life contexts ensuring natural conversational flow and emotionally expressive delivery.
- Professional quality: Recorded in acoustically treated neutral studios using high-end cameras and microphones.
- Novel collection framework: Designed to minimize equipment interference, enabling authentic, nuanced capture of both acoustic and visual emotional signals.
Data Format
Data info
dataset_root/
├── raw_videos # Original Channel Separation video recordings,
│── video_data_part1.tar.gz
│── video_data_part2.tar.gz
│── video_data_part3.tar.gz
├── {model_group}_{video_i_md5}_{left.mp4}
├── {model_group}_{video_i_md5}_{right.mp4}
└── ...
│── EmoDialogCN_prompt.json # Dialogue Prompt
│── EmoDialogCN_prompt_cn.json # Chinese Dialogue Prompt
└── processed_samples/ # Extracted conversational segments
├── input.mp4
├── wav_left_segments # Left speaker diarization result
├── wav_right_segments # Right speaker diarization result
├── test_left_speaker_diarization_asr.json # Left speaker ASR result
├── test_right_speaker_diarization_asr.json # Right speaker ASR result
├── video_segment # Dialogue Clip Generation, the original video is split into short clips organized by dialogue turns, facilitating downstream multimodal analysis.
│ ├── 000000.mp4
│ ├── 000001.mp4
│ ├── 000002.mp4
│ └── ...
└── ...
Intended Uses
- ✅ Suitable for: machine learning training, statistical analysis, visualization research
- ❌ Not suitable for: direct commercial use (specify limitations)