awakening-ai/ReactNet

ResponseNet

ResponseNet is a large-scale dyadic video dataset designed for Online Multimodal Conversational Response Generation (OMCRG). It fills the gap left by existing datasets by providing high-resolution, split-screen recordings of both speaker and listener, separate audio channels, and word‑level textual annotations for both participants.

Paper

If you use this dataset, please cite:

ResponseNet: A High‑Resolution Dyadic Video Dataset for Online Multimodal Conversational Response Generation
Authors: Luo, Cheng and Wang, Jianghui and Li, Bing and Song, Siyang and Ghanem, Bernard

Github Project

Features

• 696 temporally synchronized dyadic video pairs (over 14 hours total).
• High-resolution (1024×1024) frontal‑face streams for both speaker and listener.
• Separate audio channels for fine‑grained verbal and nonverbal analysis.
• Word‑level textual annotations for both participants.
• Longer clips (average 73.39 s) than REACT2024 (30 s) and Vico (9 s), capturing richer conversational exchanges.
• Diverse topics: professional discussions, emotionally driven interactions, educational settings, interdisciplinary expert talks.
• Balanced splits: training, validation, and test sets with equal distributions of topics, speaker identities, and recording conditions.

Data Fields

Each example in the dataset is a dictionary with the following fields:

• video/speaker: Path to the speaker’s video stream (1024×1024, frontal view).
• video/listener: Path to the listener’s video stream (1024×1024, frontal view).
• audio_speaker: Path to the speaker’s separated audio channel.
• audio/listener: Path to the listener’s separated audio channel.
• transcript/speaker: Word‑level transcription for the speaker (timestamps included).
• transcript/listener: Word‑level transcription for the listener (timestamps included).
• vector/speaker: Path to the speaker’s facial attributes.
• vector/listener: Path to the listener’s facial attributes.

Dataset Splits

We follow a standard 6:2:2 split ratio, ensuring balanced distributions of topics, identities, and recording conditions:

Split	# Video Pairs	Proportion (%)
Train	417	59.9
Valid	139	20.0
Test	140	20.1
Total	696	100.0

Visualization

You can visualize word‑cloud statistics, clip‑duration distributions, and topic breakdowns using standard Python plotting tools.

Citation

@article{luo2025omniresponse,
  title={OmniResponse: Online Multimodal Conversational Response Generation in Dyadic Interactions},
  author={Luo, Cheng and Wang, Jianghui and Li, Bing and Song, Siyang and Ghanem, Bernard},
  journal={arXiv preprint arXiv:2505.21724},
  year={2025}
}}

License

This dataset is released under the CC BY-NC 4.0 license.