Alexandre Rouxel (EBU Technology & Innovation)
Among EBU Members, there is a continued and increasing focus on how AI technology can be applied to better serve audiences in line with the public service remit. The EBU has created a new platform, the AI Hub, that facilitates the development and evaluation of customized AI solutions based on open-source models.
The idea is to facilitate a collaborative ecosystem that enables AI and media experts to co-create and refine models tailored specifically for media applications. The AI Hub, developed by the EBU’s AIM (Artificial Intelligence and Metadata) group, leverages open-source AI models, an extremely active domain.
Open-source models can play a crucial role in allowing Members to adapt to rapidly evolving technological landscapes. Such models serve as robust starting points that can be fine-tuned for tasks ranging from content analysis to content creation, addressing the diverse needs of media applications.
Importantly, the AI Hub provides private spaces for testing AI models with proprietary content. This guarantees comprehensive evaluations across a wide range of media content.
The AI Hub is hosted in the cloud and incorporates a mechanism for evaluating AI tools based on user-driven feedback. It also provides a discussion forum to facilitate collaboration between AI users and designers.
Three AI projects
Initially, the hub is hosting three projects, focused on enriching radio experiences, face recognition for television programmes, and fake news detection.
The MetaRadio project enhances radio programmes with rich metadata with a view to transforming the user experience. It uses advanced transcription and ‘diarization’ technologies to enable multi-dimensional radio experiences with improved navigation and content discoverability.
This project arose from an EBU collaboration with Radio France, who supported the development of an annotated dataset that was used for work on detecting key questions in radio programmes.
MetaRadio uses translations provided by the EBU’s EuroVOX in its processing workflow. This component is used for mood detection or as a way to address minority languages that are not well supported by AI tools.
Designed specifically for television content and with archival staff in mind, the facial recognition system prioritizes identifying active participants while excluding incidental characters. It implements face- selection algorithms to focus on relevant faces in, say, a political debate while ignoring audience members who might appear in the background.
Respect for privacy regulations is an important consideration for PSM organizations; the system operates in a way that ensures compliance with GDPR.
This system is the output of a collaboration between the EBU, Rai (Italy) and RTS (Switzerland). It uses a unique dataset of annotated video developed by the EBU based on content provided by the partners already mentioned, along with BBC and France Télévisions.
The third tool included in the AI Hub at launch is a fake news analyser. It uses natural language processing and machine learning to analyse the linguistic properties of news articles and detect potentially misleading information. The tool assigns reliability scores based on lexical, grammatical and semantic features.
These three tools are all good examples of how AI can enhance the richness and accessibility of media content. EBU Members are invited to take advantage of the AI Hub for knowledge sharing and collaboration. With the three tools above, and others that will be added over time, it is hoped that the Hub will facilitate faster adoption of state-of-the-art AI technologies in a responsible way.
As the media industry continues to evolve, embracing open-source AI will be essential for maintaining competitiveness and driving future growth.
EBU Members who want to get involved should join the AIM group: https://tech.ebu.ch/groups/aim
This article was first published in the September 2024 issue of tech-i magazine.