90,000 hours of media archive enhanced with detailed metadata

2 Apr 2025

Watching or listening to 90,000 hours of audiovisual recordings would take you over 10 years. And then they’re still not described! To prepare these valuable pieces of the past for the future more swiftly, we therefore teamed up with VRT and regional broadcasters in the Shared AI project. We used artificial intelligence to help us provide more and uniform metadata across media players – greatly improving the searchability of the archives. A good year and a lot of new metadata later, now we’re here with the results!

*On VRT’s material, we only applied entity and speech recognition. A smaller set was processed because a large part of the VRT archive had already been enriched with metadata in the past.

Shared AI is not the first time we’ve done any AI experimentation. We first dipped our toes into a smart application 10 years ago, and we’ve completed several more projects since then. Shared AI was a logical follow-up to the GIVE project, where we worked on speech, entity and face recognition for cultural and government organisations. In Shared AI, we bridged the gap to media – also making their archive material more findable with these additional descriptions.

This project has helped us unlock our archive, which had little to no annotation in its first 10 years. We can now navigate our early video material much more easily!

- Dirk Roelens, editor-in-chief at ROB tv

The power of collaboration

What’s special about this project is that the broadcasters also shared their sources alongside the algorithms and economies of scale created by combining large amounts of material. What this means in practice: suppose person X appears in 34 videos from 12 different organisations. When one of those organisations names person X, the other partners also receive new metadata. This results in links between material from different organisations, more uniformity, and time savings.

To make the material even more searchable, the new descriptions were also linked to the external source, Wikidata. VRT held the reins in this process.

Looking to the future

Artificial intelligence is very promising for annotating large volumes of data – and with this project we’ve demonstrated this for the broadcasters!

- Matthias Priem, Manager Archivering at meemoo

It would be a shame to stop the story with these great results – and that's why part of the work will continue: broadcasters can keep linking the most frequently occurring unknown persons in their archive to a name, thus adding extra metadata. We’re also taking our insights to the Hermes project that started last year. What’s the plan, and where are we at?

Want to dig deeper into the learnings from previous AI projects and our further ambitions around artificial intelligence? We summarised it on the knowledge database (article in Dutch).

This project was part of the Flemish Government’s Resilience Recovery Plan.

This page is loading...