Artificial Intelligence and audiovisual archives: lessons learned

Abstract

This lightning talk will highlight the results and lessons learned from the GIVE project. In this project, meemoo processed 170.000 hours of audiovisual archives using speech-to-text, named entity recognition and face recognition. We created time-based metadata identifying people, locations and organisations in audio and video. Where possible, metadata are linked to open linked authorities, such as wikidata. Over 120 partners were involved in the project, from various sectors such as digital heritage, city archives, museums and political archives. Part of this talk will be dedicated to how we involved these partners into the new but also very challenging field of creating metadata through AI. The lightning talk is not intended as a deep dive into technical matters. We will highlight the obtained results and discuss the quality of obtained result. But more importantly, we will explain why and how we decided to build certain components, such as face recognition, while opting to procure speech-to-text and NER services. This should give the audience an idea of the decision processes behind working with AI tooling to create descriptive metadata. We will also briefly touch the ethical and legal challenges we encountered, share insights that came out of the project. Lastly we will end the talk by looking at the future and share the questions that we still have and need to tackle next. We hope this will provide inspiration for discussions with the iPRES community during the conference.

Details

Creators

Matthias Priem

Institutions

Date

2024-09-18 11:20:00 +0100

Keywords

metadata standards and implementation; scaling up

Publication Type

lightning talk

License

Creative Commons Zero (CC0-1.0)

Download

(unknown) bytes

Slides

here

Video Stream

here

Collaborative Notes

here

View This Publication