---
language:
- en
tags:
- webdataset
- audio
- motion
---

## UniLS-Talk Dataset

To enable research on unified speaking and listening avatar generation, we curate and construct the **UniLS-Talk Dataset**, a large-scale collection of high-quality 3D facial motion data. We apply a carefully designed tracking pipeline to extract per-frame [FLAME](https://flame.is.tue.mpg.de/) parameters, including expression coefficients, eye-gaze, jaw pose and head pose annotations. The dataset comprises two complementary parts:

- **Paired conversational data** sourced from the [Seamless Interaction](https://ai.meta.com/research/seamless-interaction/) dataset, providing synchronized dual-speaker videos with natural turn-taking dynamics between speaking and listening.
- **Unpaired multi-scenario data** aggregated from CelebV, TalkingHead-1KH, TEDTalk, VFHQ, and other in-the-wild videos, covering diverse facial behaviors across identities and environments (news broadcasts, interviews, casual talking, etc.).

| Category | Source | Hours | Audio | Motion |
|----------|--------|-------|-------|--------|
| **Paired Conversational** | Seamless Interaction Dataset | 657.5 h | ✅ | ✅ |
| **Unpaired Multi-Scenario** | Diverse identities and environments from in-the-wild videos | **546.5 h** | ❌ | ✅ |
| **Total** | | **1,204 h** | | |

The paired conversational data is split into **622.5 hours** for training, **4.8 hours** for validation, and **30.2 hours** for testing. All data includes FLAME expression parameters, jaw and head pose, and eye-gaze annotations at 25 fps.