StoryTeller: Training-Free Narrative Grounding for Long-Form Audio Description
StoryTeller is a training-free framework for long-form audio description that maintains a verified narrative memory to carry story-relevant information across scenes, using only raw video and a movie title, with optional retrieval of public movie metadata. It requires no subtitles, scripts, AD transcripts, aligned captions, character banks, precomputed face identities, or task-specific fine-tuning.
StoryTeller's approach of maintaining a verified narrative memory without training suggests a potential direction for improving long-form video understanding in VLMs. The next signal to watch is whether this method can be integrated into existing VLMs to enhance narrative coherence in tasks beyond audio description.
This work addresses a key limitation of current VLMs in handling long-form narrative context, which is relevant for accessibility applications. The next signal is adoption by accessibility tool developers or integration into video streaming platforms.
StoryTeller could enable more accessible long-form video content for blind and low-vision audiences without requiring expensive fine-tuning or additional data. The next signal is interest from accessibility-focused companies or video platforms.
If StoryTeller proves effective, it could lead to training-free methods for maintaining narrative context in other long-form video tasks. The next signal is follow-up work applying similar memory mechanisms to video question answering or summarization.