DualTalk Dataset
π Overview
The DualTalk dataset is the first large-scale dataset specifically designed for dual-speaker 3D talking head conversation generation. This dataset supports speaker and listener role transitions, multi-round conversations, and natural interactions, providing essential benchmark data for the 3D talking head generation field.
π― Dataset Features
- Dual-Speaker Interaction: Contains synchronized audio and facial expression data from two speakers
- Role Transition: Supports dynamic transitions between speaker and listener roles
- Multi-Round Conversations: Provides continuous multi-round conversation data
- High-Quality Annotations: Includes precise FLAME parameters and audio features
- Diverse Scenarios: Covers various conversation scenarios and emotional expressions
π Dataset Structure
DualTalk_Dataset/
βββ train/
β βββ xxx_speaker1.wav # Speaker 1 audio files
β βββ xxx_speaker1.npz # Speaker 1 FLAME parameters
β βββ xxx_speaker2.wav # Speaker 2 audio files
β βββ xxx_speaker2.npz # Speaker 2 FLAME parameters
β βββ ...
βββ test/
β βββ xxx_speaker1.wav
β βββ xxx_speaker1.npz
β βββ xxx_speaker2.wav
β βββ xxx_speaker2.npz
β βββ ...
βββ ood/ # Out-of-distribution test data
βββ xxx_speaker1.wav
βββ xxx_speaker1.npz
βββ xxx_speaker2.wav
βββ xxx_speaker2.npz
βββ ...
π Citation
If you use the DualTalk dataset, please cite our paper:
@inproceedings{peng2025dualtalk,
title={Dualtalk: Dual-speaker interaction for 3d talking head conversations},
author={Peng, Ziqiao and Fan, Yanbo and Wu, Haoyu and Wang, Xuan and Liu, Hongyan and He, Jun and Fan, Zhaoxin},
booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
pages={21055--21064},
year={2025}
}
π License
This dataset is derived from the RealTalk dataset and is licensed under Apache 2.0.
Apache 2.0 License: This license allows you to use, modify, and distribute the dataset for both commercial and non-commercial purposes, with proper attribution and license inclusion.
For the complete license text, please see: Apache 2.0 License