--- license: cc-by-4.0 task_categories: - text-to-3d - robotics - text-to-speech language: - zh tags: - interactive - motion - avatar size_categories: - 10K arXiv License Clips Duration

A Large-Scale Multimodal Dialogue Corpus with Synchronized Speech, Full-Body Motion, and Facial Expressions

From the paper: SentiAvatar: Towards Expressive and Interactive Digital Humans

--- ## Overview **SuSuInterActs** is a high-quality dialogue motion capture dataset built around a single virtual character **SuSu (่‹่‹)**. The dataset features synchronized multi-modal data captured via professional optical motion capture, including: - ๐Ÿ—ฃ๏ธ **Speech Audio** โ€” Natural Chinese conversational speech - ๐Ÿ’ƒ **Full-Body Motion** โ€” 63-joint skeleton with body, hands, and fingers (6D rotation) - ๐ŸŽญ **Facial Expressions** โ€” 51-dimensional ARKit BlendShape coefficients - ๐Ÿ“ **Rich Text Annotations** โ€” Action tags, expression tags, and dialogue transcripts

Dataset Modality Overview

### Key Statistics | Stat | Value | |------|-------| | **Total clips** | 21,133 | | **Total duration** | ~37 hours | | **Avg. clip duration** | ~5.4 seconds | | **Frame rate** | 20 FPS (motion & face), 16kHz (audio) | | **Skeleton** | 63 joints (25 body + 20 left hand + 20 right hand) | | **Face dims** | 51 (ARKit BlendShape) | | **Language** | Chinese (Mandarin) | | **Splits** | Train: 19,019 / Val: 635 / Test: 1,479 | ## ๐Ÿ“ Directory Structure ``` SuSuInterActs/ โ”œโ”€โ”€ README.md โ”œโ”€โ”€ assets/ # Figures for this README โ”‚ โ”œโ”€โ”€ motion_data/ # ๐Ÿ’ƒ Full-body motion data (4.9 GB) โ”‚ โ”œโ”€โ”€ fbx_to_json_data_susu_retarget_maya/ โ”‚ โ”‚ โ”œโ”€โ”€ 20250801/ โ”‚ โ”‚ โ”‚ โ”œโ”€โ”€ Human_xxx.npy โ”‚ โ”‚ โ”‚ โ””โ”€โ”€ ... โ”‚ โ”‚ โ”œโ”€โ”€ 20250804/ โ”‚ โ”‚ โ””โ”€โ”€ ... (40+ capture sessions) โ”‚ โ””โ”€โ”€ fbx_to_json_data_susu_chonglu/ โ”‚ โ”œโ”€โ”€ 20260115/ โ”‚ โ””โ”€โ”€ ... โ”‚ โ”œโ”€โ”€ wav_data/ # ๐Ÿ”Š Speech audio (6.3 GB) โ”‚ โ”œโ”€โ”€ fbx_to_json_data_susu_retarget_maya/ โ”‚ โ”‚ โ””โ”€โ”€ ... (same structure as motion_data) โ”‚ โ””โ”€โ”€ fbx_to_json_data_susu_chonglu/ โ”‚ โ””โ”€โ”€ ... โ”‚ โ”œโ”€โ”€ arkit_data/ # ๐ŸŽญ Facial expression data (750 MB) โ”‚ โ””โ”€โ”€ fbx_to_json_data_susu_retarget_maya/ โ”‚ โ””โ”€โ”€ ... (same structure) โ”‚ โ”œโ”€โ”€ text_data/ # ๐Ÿ“ Text annotations (8 MB) โ”‚ โ”œโ”€โ”€ motion2text.json # Main annotation file: name โ†’ text+tags โ”‚ โ”œโ”€โ”€ train.json # Training set annotations โ”‚ โ”œโ”€โ”€ val.json # Validation set annotations โ”‚ โ””โ”€โ”€ test.json # Test set annotations โ”‚ โ””โ”€โ”€ split/ # ๐Ÿ“‹ Data splits โ”œโ”€โ”€ all_file_list.txt # 21,133 entries โ”œโ”€โ”€ train_file_list.txt # 19,019 entries โ”œโ”€โ”€ val_file_list.txt # 635 entries โ””โ”€โ”€ test_file_list.txt # 1,479 entries ``` **Total size: ~12 GB** ## ๐Ÿ“Š Data Formats ### Motion Data (`motion_data/*.npy`) Each `.npy` file stores a Python dictionary with 4 keys: ```python import numpy as np data = np.load("motion_data/fbx_to_json_data_susu_retarget_maya/20250801/Human_xxx.npy", allow_pickle=True).item() data["body"] # (T, 153) โ€” root offset velocity (3) + body 6D rotation (25ร—6) data["left"] # (T, 120) โ€” left hand 6D rotation (20ร—6) data["right"] # (T, 120) โ€” right hand 6D rotation (20ร—6) data["positions"] # (T, 63, 3) โ€” 3D joint positions (for visualization) ``` | Key | Shape | Description | |-----|-------|-------------| | `body` | `(T, 153)` | Root offset velocity (3D) + 25 body joints ร— 6D rotation | | `left` | `(T, 120)` | 20 left hand joints ร— 6D rotation | | `right` | `(T, 120)` | 20 right hand joints ร— 6D rotation | | `positions` | `(T, 63, 3)` | Global 3D joint positions (63 joints ร— xyz) | - **Frame rate**: 20 FPS - **Rotation representation**: 6D rotation - **Root displacement**: The first 3 dims of `body` encode root translation velocity (differential encoding). To recover absolute position, accumulate: `pos[t] = pos[t-1] + vel[t]` ### 63-Joint Skeleton ``` Body (25 joints): pelvis โ†’ thigh_r โ†’ calf_r โ†’ foot_r โ†’ ball_r โ†’ thigh_l โ†’ calf_l โ†’ foot_l โ†’ ball_l โ†’ spine_01 โ†’ spine_02 โ†’ spine_03 โ†’ spine_04 โ†’ spine_05 โ†’ neck_01 โ†’ neck_02 โ†’ head โ†’ clavicle_l โ†’ upperarm_l โ†’ lowerarm_l โ†’ hand_l โ†’ clavicle_r โ†’ upperarm_r โ†’ lowerarm_r โ†’ hand_r Left Hand (20 joints): hand_l โ†’ index[0-3] โ†’ middle[0-3] โ†’ ring[0-3] โ†’ pinky[0-3] โ†’ thumb[0-2] Right Hand (20 joints): hand_r โ†’ index[0-3] โ†’ middle[0-3] โ†’ ring[0-3] โ†’ pinky[0-3] โ†’ thumb[0-2] ``` ### Facial Data (`arkit_data/*.npy`) ```python face = np.load("arkit_data/.../Human_xxx.npy") # shape: (T, 51), dtype: float64 ``` 51-dimensional ARKit BlendShape coefficients (values in [0, 1]): | Index | BlendShape | Index | BlendShape | |-------|-----------|-------|-----------| | 0 | browDownLeft | 26 | mouthClose | | 1 | browDownRight | 27 | mouthDimpleLeft | | 2 | browInnerUp | 28 | mouthDimpleRight | | 3 | browOuterUpLeft | ... | ... | | 8 | eyeBlinkLeft | 43 | mouthSmileLeft | | 9 | eyeBlinkRight | 44 | mouthSmileRight | | 24 | jawOpen | 50 | noseSneerRight | ### Audio Data (`wav_data/*.wav`) - **Format**: WAV, 16-bit PCM - **Sample Rate**: 16,000 Hz (mono) - **Language**: Chinese (Mandarin) ### Text Annotations (`text_data/motion2text.json`) Each entry maps a clip name to its annotation string: ```json { "fbx_to_json_data_susu_retarget_maya/20250826/Human_0825_153-5_01": "ใ€่กจๆƒ…๏ผšๅพฎ็ฌ‘่ฏข้—ฎใ€‘ใ€ๅŠจไฝœ๏ผšๅคดๅพฎๅ‘ๅณๆญชใ€‘่ฟ˜ๆœ‰็กๅ‰ๅ‡†ๅค‡ๅ•ฅ็š„...", "fbx_to_json_data_susu_chonglu/20260115/Human_100_73_01_B": "ใ€่กจๆƒ…๏ผš็œผ็ฅž่ฎค็œŸใ€‘ใ€ๅŠจไฝœ๏ผš่บซไฝ“ๅพฎๅ‰ๅ€พใ€‘ๅฎ‰ๅฎ‰๏ผŒไฝ ่ทŸๅงๅง่ฏดๅฎž่ฏใ€‚" } ``` **Annotation format**: `ใ€่กจๆƒ…๏ผšใ€‘ใ€ๅŠจไฝœ๏ผšใ€‘` - **Expression tags** (่กจๆƒ…): e.g., ๅพฎ็ฌ‘ (smile), ่ฎค็œŸ (serious), ๆ‹…ๅฟง (worried), ่ฐƒ็šฎ (playful) - **Action tags** (ๅŠจไฝœ): e.g., ็ผ“ๆ…ข็‚นๅคด (slow nod), ๅŒ่‡‚ๅฑ•ๅผ€ (arms spread), ๅคดๅพฎๅ‘ๅณๆญช (head tilt right) ### Split Files (`split/*.txt`) Each line is a relative path (without extension) identifying a clip: ``` fbx_to_json_data_susu_chonglu/20260115/Human_82_84_01_B fbx_to_json_data_susu_retarget_maya/20250905/Human_0904_152-8_01 ... ``` Use these to load the corresponding files: ```python name = "fbx_to_json_data_susu_retarget_maya/20250905/Human_0904_152-8_01" motion = np.load(f"motion_data/{name}.npy", allow_pickle=True).item() face = np.load(f"arkit_data/{name}.npy") audio = f"wav_data/{name}.wav" text = motion2text[name] ``` ## ๐Ÿ”ง Quick Start ### Load a sample ```python import numpy as np import json import soundfile as sf # Load split with open("split/test_file_list.txt") as f: test_names = [line.strip() for line in f if line.strip()] name = test_names[0] # Load motion motion = np.load(f"motion_data/{name}.npy", allow_pickle=True).item() print(f"Body: {motion['body'].shape}") # (T, 153) print(f"Hands: {motion['left'].shape}") # (T, 120) print(f"Positions: {motion['positions'].shape}") # (T, 63, 3) # Load face face = np.load(f"arkit_data/{name}.npy") print(f"Face: {face.shape}") # (T_face, 51) # Load audio audio, sr = sf.read(f"wav_data/{name}.wav") print(f"Audio: {audio.shape}, sr={sr}") # Load text annotation with open("text_data/motion2text.json") as f: motion2text = json.load(f) print(f"Text: {motion2text[name]}") ``` ### Convert to BVH (for visualization) See [SentiAvatar](https://github.com/yourorg/SentiAvatar) for the visualization tool: ```bash python tools/visualize_motion.py \ --input SuSuInterActs/motion_data/path/to/sample.npy \ --output output.bvh ``` ## ๐Ÿ“Š Data Distribution

Duration and Text Distribution

## ๐Ÿ“ Citation If you use this dataset in your research, please cite: ```bibtex @article{jin2026sentiavatar, title={SentiAvatar: Towards Expressive and Interactive Digital Humans}, author={Jin, Chuhao and Zhang, Rui and Gao, Qingzhe and Shi, Haoyu and Wu, Dayu and Jiang, Yichen and Wu, Yihan and Song, Ruihua}, journal={arXiv preprint arXiv:2604.02908}, year={2026} } ``` ## License This dataset is released under [CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International)](../SentiAvatar/LICENSE). - โœ… Free for academic and non-commercial research - โŒ Not permitted for commercial use - ๐Ÿ“ง Contact the authors for commercial licensing ## Acknowledgments This dataset was captured at SentiPulse using professional optical motion capture equipment. We thank all participants and the annotation team for their contributions.