arXiv preprint 2025

ALIVE

ALIVE: An Avatar-Lecture Interactive Video Engine with Content-Aware Retrieval for Real-Time Interaction

Md Zabirul Islam, Md Motaleb Hossen Manik, Ge Wang

Current or former lab member. Select a name to see all of their projects.

Method overview of ALIVE
System overview of ALIVE (Fig. 1 of the paper).

Abstract

Recorded lectures are flexible but give students no way to ask for clarification on the spot. ALIVE (Avatar-Lecture Interactive Video Engine) adds that interaction and runs entirely on local hardware. It generates avatar-delivered lecture segments from speech transcription, language-model refinement and talking-head synthesis.

Content-aware retrieval grounds each answer in the lecture material. Students can pause the video, ask by text or voice, and receive a concise answer as text or from the avatar. The system was evaluated on a complete medical imaging course for accuracy, latency and user experience.

What’s in the repository

  • The deployed backend: retrieval, answer generation, and audio and avatar routing
  • Scripts for building the retrieval index, LoRA fine-tuning, and the question-answer data pipeline
  • Evaluation and latency benchmarks with the result summaries used in the paper

Citation

@article{islam2025alive,
  title   = {ALIVE: An Avatar-Lecture Interactive Video Engine with Content-Aware Retrieval for Real-Time Interaction},
  author  = {Islam, Md Zabirul and Manik, Md Motaleb Hossen and Wang, Ge},
  journal = {arXiv preprint arXiv:2512.20858},
  year    = {2025}
}
Back to projects