SuSuInterActs Dataset

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 (苏苏). The dataset features synchronized multi-modal data captured via professional optical motion capture, including:

• 🗣️ Speech Audio — Natural Chinese conversational speech
• 💃 Full-Body Motion — 63-joint skeleton with body, hands, and fingers (6D rotation)
• 🎭 Facial Expressions — 51-dimensional ARKit BlendShape coefficients
• 📝 Rich Text Annotations — Action tags, expression tags, and dialogue transcripts

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

📁 Directory Structure

SuSuInterActs/
├── README.md
├── assets/                          # Figures for this README
│
├── motion_data/                     # 💃 Full-body motion data (4.9 GB)
│   ├── fbx_to_json_data_susu_retarget_maya/
│   │   ├── 20250801/
│   │   │   ├── Human_xxx.npy
│   │   │   └── ...
│   │   ├── 20250804/
│   │   └── ...  (40+ capture sessions)
│   └── fbx_to_json_data_susu_chonglu/
│       ├── 20260115/
│       └── ...
│
├── wav_data/                        # 🔊 Speech audio (6.3 GB)
│   ├── fbx_to_json_data_susu_retarget_maya/
│   │   └── ... (same structure as motion_data)
│   └── fbx_to_json_data_susu_chonglu/
│       └── ...
│
├── arkit_data/                      # 🎭 Facial expression data (750 MB)
│   └── fbx_to_json_data_susu_retarget_maya/
│       └── ... (same structure)
│
├── text_data/                       # 📝 Text annotations (8 MB)
│   ├── motion2text.json             # Main annotation file: name → text+tags
│   ├── train.json                   # Training set annotations
│   ├── val.json                     # Validation set annotations
│   └── test.json                    # Test set annotations
│
└── split/                           # 📋 Data splits
    ├── all_file_list.txt            # 21,133 entries
    ├── train_file_list.txt          # 19,019 entries
    ├── val_file_list.txt            #    635 entries
    └── test_file_list.txt           #  1,479 entries

Total size: ~12 GB

📊 Data Formats

Motion Data (motion_data/*.npy)

Each .npy file stores a Python dictionary with 4 keys:

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)  — root offset velocity (3) + body 6D rotation (25×6)
data["left"]       # (T, 120)  — left hand 6D rotation (20×6)
data["right"]      # (T, 120)  — right hand 6D rotation (20×6)
data["positions"]  # (T, 63, 3) — 3D joint positions (for visualization)

Key	Shape	Description
body	(T, 153)	Root offset velocity (3D) + 25 body joints × 6D rotation
left	(T, 120)	20 left hand joints × 6D rotation
right	(T, 120)	20 right hand joints × 6D rotation
positions	(T, 63, 3)	Global 3D joint positions (63 joints × 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 → thigh_r → calf_r → foot_r → ball_r
         → thigh_l → calf_l → foot_l → ball_l
         → spine_01 → spine_02 → spine_03 → spine_04 → spine_05
           → neck_01 → neck_02 → head
           → clavicle_l → upperarm_l → lowerarm_l → hand_l
           → clavicle_r → upperarm_r → lowerarm_r → hand_r

Left Hand (20 joints):
  hand_l → index[0-3] → middle[0-3] → ring[0-3] → pinky[0-3] → thumb[0-2]

Right Hand (20 joints):
  hand_r → index[0-3] → middle[0-3] → ring[0-3] → pinky[0-3] → thumb[0-2]

Facial Data (arkit_data/*.npy)

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:

{
  "fbx_to_json_data_susu_retarget_maya/20250826/Human_0825_153-5_01": 
    "【表情：微笑询问】【动作：头微向右歪】还有睡前准备啥的...",
  
  "fbx_to_json_data_susu_chonglu/20260115/Human_100_73_01_B": 
    "【表情：眼神认真】【动作：身体微前倾】安安，你跟姐姐说实话。"
}

Annotation format: 【表情：<expression_tag>】【动作：<action_tag>】<dialogue_transcript>

• Expression tags (表情): e.g., 微笑 (smile), 认真 (serious), 担忧 (worried), 调皮 (playful)
• Action tags (动作): e.g., 缓慢点头 (slow nod), 双臂展开 (arms spread), 头微向右歪 (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:

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]

🔧 Quick Start

Load a sample

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 for the visualization tool:

python tools/visualize_motion.py \
    --input SuSuInterActs/motion_data/path/to/sample.npy \
    --output output.bvh

📊 Data Distribution

📝 Citation

If you use this dataset in your research, please cite:

@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).

• ✅ Free for academic and non-commercial research
• ❌ Not permitted for commercial use
• 📧 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.