WildVid-LIP: In-The-Wild Temporal Anchors for Visual Speech Recognition

WildVid-LIP is a large-scale, open-source dataset mapping over 100,000 curated temporal segments from unconstrained, real-world YouTube videos. It provides precise timestamp anchors optimized for training Visual Speech Recognition (VSR / Lip-Reading), audio-visual synchronization, and multimodal self-supervised models.

Instead of distributing heavy, monolithic video files—which introduces platform friction and restrictive redistribution licensing—WildVid-LIP distributes high-fidelity metadata anchors. This allows researchers to stream or crop localized face tracks directly from source media while maintaining absolute compliance with digital copyright frameworks.

---

Key Features

• Real-World Complexity: Unlike laboratory-controlled data (e.g., GRID, TCD-TIMIT), these anchors capture dynamic variations in head pose, natural lighting transitions, speaker diversity, compression artifacts, and occlusions.
• Scale-Optimized: Expanded to 100,000+ precise windows, making it ideal for deep spatial-temporal networks (3D-CNNs, Vision Transformers, and Conformers).
• Self-Supervised & Multimodal Ready: Designed intentionally without rigid text baselines, enabling pipelines to dynamically generate pseudo-labels using state-of-the-art Automated Speech Recognition (ASK) models like Whisper or Wav2Vec2 during ingestion.

---

Dataset Structure

The corpus is contained within a single manifest mapping (train.csv). The schema tracks highly accurate temporal bounds across active speech windows:

Column Name	Data Type	Description
video_name	string	The unique 11-character YouTube video alphanumeric identifier.
time_start	float	The precise timestamp (in seconds) where the targeted visual speech sequence begins.
time_end	float	The precise timestamp (in seconds) where the targeted visual speech sequence terminates.

Sample Schema Entry

video_name,time_start,time_end
dQw4w9WgXcQ,42.15,44.50

Future Updates

WildVid-Lip would be regularly updated with transcripts, download scripts, preprocessed videos for direct downloads and much more!