A major television broadcaster holds tens of thousands of programs accumulated over decades. Older footage often lacks detailed metadata, so locating particular scenes can depend on the memory of experienced archive staff.
Technica AI evaluated AI-generated scene descriptions, keywords, and summaries using video, audio, and text. A five-program proof of concept assessed the feasibility of content-based discovery, providing evidence for further development of archive discovery and potential licensing services.
A large archive with limited content-level metadata.
The broadcaster’s archive spans decades of television history and includes tens of thousands of programs. The company also provides selected archive material to third parties.
Older records often contain little more than a title and broadcast date. Without information about individual scenes, valuable footage is difficult to locate or assess for potential reuse.
Search requests rarely arrive as exact program titles.
Editors and external users may need footage of a particular event, a person speaking, or a scene conveying a particular setting or mood.
Existing search fields could not reliably answer those requests. Staff had to rely on memory and manually preview candidate programs, limiting the usefulness of the archive.
Creating searchable metadata from video, audio, and text.
Technica AI analyzed footage to generate candidate scene boundaries, keywords, descriptions, summaries, and representative images.
The PoC used five programs across news, information, drama, variety, and documentary genres to assess whether generated metadata could support practical content discovery.
The objective was to search for people, topics, events, places, spoken content, and scene context—not only titles and broadcast dates.
PoC accuracy varied across metadata tasks and programs.
The five-program PoC evaluated broadcasts dating from 1959 to 1996. The reported accuracy ranges are shown below. The figures apply to the selected material, with broader archive coverage requiring further evaluation.
| Metric | PoC result range |
|---|---|
| Keyword accuracy | 75.6%–96.6% |
| Scene description accuracy | 48.57%–88.89% |
| Scene summary accuracy | 61.9%–88.89% |
In an additional documentary evaluation, 400 of 429 extracted keywords matched the footage, equivalent to approximately 93.24%. This is a separate observation from the five-program accuracy ranges.
The results established initial potential for AI-assisted discovery, including older footage, while showing that description and summarization quality varied substantially between tests.
A clearer route from archive storage to content reuse.
-
More targeted archive search
Content-level metadata can help staff identify candidate programs and scenes.
-
Potential reuse of older footage
Material with sparse original metadata becomes easier to discover and review.
-
A basis for new services
Searchable scenes may support licensing, thematic collections, and content offerings.
-
Reduced dependence on memory
Searchable descriptions can complement the experience of long-serving archive staff.
Making archived video easier to discover and evaluate.
The project demonstrated a path toward content-based archive search and future licensing workflows. Scaling the approach would require further evaluation across a broader range of programs and production requirements.