Evaluating and enhancing AI-generated audio description for factual television content
Description
As demand for accessible media grows, interest in automating audio description (AD) with generative AI has accelerated. Yet AD is a complex practice in which quality is crucial to meaningful access. This presentation reports a two-phase study, commissioned by the Royal National Institute of Blind People (RNIB) and conducted in 2025–26, combining comparative evaluation, prompt engineering and user-centred assessment to examine the feasibility of AI-generated AD for factual television.
The first study compared zero-shot AD from recent vision-language and large language models on factual BBC programmes, identifying persistent limitations in character tracking, accuracy, saliency and cohesion, alongside user and expert concerns about trust in automation. The second examined the extent to which structured prompting could improve AI outputs.
User feedback broadly confirmed improvements but highlighted persistent variability, pacing problems and inaccuracies, as well as differing preferences among users. In light of these findings, an important point for discussion will therefore be the role of human expertise in AI-enabled approaches to AD generation.
Hosted by Professor Andrea Rizzi and Dr Yu Hao, School of Languages and Linguistics, Faculty of Arts
Lineup
Professor Sabine Braun
Dr Sabine Braun is Professor of Translation Studies, Director of the Centre for Translation Studies at the University of Surrey, and a Co-Director of the Surrey Institute for People-Centred AI. Her research explores human-machine interaction and integration in cross-lingual and cross-modal mediation (e.g., interpreting, AD) to improve access to information, media, and public services. She has a long-standing interest in exploring the cognitive, discursive and narratological foundations of AD. In the European Horizon 2020 Project MeMAD (Methods for Managing Audiovisual Data, 2018-21), she collaborated with computer vision experts to investigate the feasibility of (semi-)automating AD creation. She co-edited ‘Innovations in Audio Description Research’ (Routledge, 2021). In 2024, she launched the Leverhulme Trust-funded Doctoral Training Network on AI-Enabled Digital Accessibility (ADA), which includes AI-enabled AD as key research strand. She has advised industry and third-sector organisations on the use of AI-generated AD.
Tickets for good, not greed Humanitix dedicates 100% of profits from booking fees to charity
Register
Register
More events from this host
