The European Parliament’s Archives Unit has developed Archibot, an innovative AI-powered tool designed to enhance access to the institution’s extensive legislative archives. By using a method known as Retrieval-Augmented Generation (RAG), Archibot combines the capabilities of large language models with authoritative archival sources to deliver accurate, context-rich responses to user queries.

Originally optimized for French-language searches, Archibot now supports multiple languages, including English, German, Spanish, Italian, Dutch, and others—reflecting the multilingual nature of the European Union. The tool enables users to perform a range of complex tasks, such as retrieving relevant documents, generating concise summaries, and receiving curated lists of archival materials.

By bridging artificial intelligence with trusted historical data, Archibot offers researchers, policymakers, and the public a powerful new way to explore the legislative history of the European Union.

The EBU AI Sandbox is a collaborative environment enabling public service media to explore, test, and refine AI tools specifically tailored to media and archive applications. By connecting AI developers with media and archive professionals, the Sandbox fosters the co-creation of user-centric applications rooted in operational realities. It offers secure, private spaces to evaluate AI using proprietary content while ensuring GDPR compliance and data protection. Grounded in an open-source philosophy, the Sandbox accelerates innovation through transparency, adaptability, and shared learning. Among its flagship projects are a GDPR-compliant facial recognition tool optimised for documentalists and MetaRadio, which transforms linear radio into a rich, searchable experience. The iterative, feedback-based process helps ensure that development efforts are focused on features that add real value to users and fit into their day-to-day workflows. This presentation illustrates how bringing users and developers together can help make AI more valuable and relevant for media archives.

At VRT, like many broadcasting organizations, the production and archiving of diverse media content has led to the existence of multiple systems with complex integrations. This often results in various “dialects” created by operational teams or the tools themselves. Archives, in particular, which import data from different systems and have their own archiving systems, suffer from different understandings and policies in the operational units. VRT Archive has been searching for a solution in the form of a “common language” to improve data exchange between systems and facilitate migrations. The EbuCorePlus ontology, which offers a semantic framework for describing processes and content in the broadcast world, seems to be the most evident standard to start with. This talk will describe our goals, planning, and experiences so far in this journey.

AI involvement in media creation and editing is now commonplace, and humans are unable to reliably discern AI generated media from that which is captured with a camera. We will examine the different places where AI and other, less advanced editing techniques have the potential to harm, and how the BBC R&D has co-developed the “Content Credentials” technical specification to help address them. We’ll cover the transparent disclosure of the “ingredients” and “actions” involved in producing media, and how we identify whether AI has been used. We’ll look at the different uses of media provenance: capture and ingest, storage and retrieval, and how consumers might make use of this extra information.

This presentation takes you through the development of automated content summaries done by the radio archive team at Yle, the Finnish Broadcasting Company. Building on our previous work in speech recognition, during 2024 we focused on creating content summaries from transcripts, aiming for high automation levels and backwards compatible metadata. We would like to share, from a practical perspective, what we have learned on topics such as comparing different LLMs and their parameters, building prompts, and deciding on what the summaries should look like. For broader context, the presentation also provides an update on current AI-related topics and plans at Yle Archives, and how projects such as this can be shared with and adopted by other teams.

The presentation will show how innovative tools are implemented into archival user interfaces at the ORF. The examples will be out of all stages between prototype and productive use. From a more generalized perspective there will be an insight how the Company learns about the potentials of AI by use of the AIditor an award winning AI-Framework available to all ORF users.

Full abstract coming soon.

Broadcast media organisations produce many news scripts every day for dissemination as content. Such text data is often reused in the process of producing TV programmes and web news. To efficiently utilise this large volume of data, it is necessary to accurately attach metadata such as topic labels that indicate the contents of the text. However, manually assigning metadata takes an enormous amount of time and effort. With the aim of reducing costs, we developed a system that automatically adds metadata to news articles. The system provides a variety of metadata, such as labels, keywords, auto-generated titles, emotion tags and speculative location when a news article is input. We focused on topic labels since they are versatile when used as metadata. We proposed a novel loss function that utilises weighting and a label smoothing technique to suppress label imbalance. Experimental results show that our method outperforms baselines.

Meemoo, the Flemish institute for archives, digitises and preserves a large amount of primarily audiovisual objects from organisations in the cultural heritage, media and government sector. More than 8 million items are currently archived, including newspapers, images, video, and audio files. Besides archiving, meemoo also tends the distribution of material to specific target groups via web platforms, including education and the general public. Behind those platforms, there’s a vast amount of metadata that is collected from our partners, which enables the archive material to be searchable and usable across the various collections.

The amount of audiovisual objects has long exceeded the available, often manually created metadata. Hence, we turn to artificial intelligence to enable the automatic annotation of audio and video at scale, which is the goal of the SharedAI project: a collaboration between meemoo, regional broadcasters and public broadcaster VRT. We were able to produce transcripts for and identify public figures and locations in numerous hours of content by applying a combination of face-recognition, speech-to-text and Named-Entity Recognition (NER).

In this presentation, we will talk about the approach and results of the SharedAI project. We also zoom in on how the output of our automatic annotation pipelines were integrated with the archive’s metadata in our Knowledge Graph, which is our uniform and application-independent access to metadata. Lastly, we will discuss how we collaboratively construct a thesaurus of public figures in Flanders targeted at face recognition software and integrate with other knowledge graphs.

Tasked by the Welsh Government to develop a foundation of Language Technology resources for the Welsh language, the Language Technologies Unit at Bangor University has developed a range of technologies aimed at facilitating the use of Welsh in digital environments. These include bilingual speech recognition models, a number of different synthetic voices and domain-specific machine translation engines as well as finetuning commercial and open source AI models. This presentation will provide an overview of the available technologies, and demonstrate their use within: Macsen, an Alexa-like app for Welsh; Trawsgrifiwr, a subtitling website; and Trosleisio, a service for adding synthesized Welsh-language voiceovers to videos. Finally, some of the challenges facing improving such models and making them available under open-source licenses will be discussed.

This presentation, co-presented by RTÉ and Dublin City University, explores a pioneering effort to digitally preserve and analyse Ireland’s rich linguistic heritage. With over a century’s worth of broadcast recordings, the project leverages advanced digitization and linguistic research, unlocking a vast dataset for scholars, educators, and the wider public. Attendees will gain insights into the methodologies, discoveries, and future applications of this vital resource, now publicly accessible via the dialect repository Canúint.ie.

The future of Gaelic, like many minority languages, is uncertain. Could language technology help secure its survival? While AI systems like ChatGPT now permit real-time English interactions, achieving this for Gaelic is challenging. This lecture examines key technological advances needed to bridge this gap and their impact on language revitalisation. Research at the University of Edinburgh explores two promising solutions to data scarcity in low-resource languages: speech recognition and synthetic data augmentation using Large Language Models. A key approach involves leveraging archival data to develop language technology and then using technology to annotate and enhance archival materials, creating a virtuous circle of resource expansion. However, technological progress also raises concerns, particularly in collaborations between major tech companies and minority language communities, where issues of data ownership and cultural preservation must be carefully managed.

How we use a LLM through an internal portal to help us to write headlines and summaries of agency items, based on their scripts. We keep the datas safe and we we save time to do work with greater added value.

Full abstract coming soon.

Sound & Vision started to experience with ChatGPT and its potential in archives management several years ago and presented about this on the World Conference in 2023. At the ingest department, experiments were conducted to summarize speech-to-text data and generate a starting point for descriptions of websites or games. Another experiment involved creating simple Python scripts using ChatGPT’s assistance for automating basic archival tasks such as MD5 checksum validation, reading media files, generating file lists, and renaming files.

What happened since those early days, did the archive embed these kind of practices in its daily practices or not?”

This talk will describe how the National Library of Wales established a National Broadcast Archive, with support from the National Lottery Heritage Fund and the Welsh Government, and is now a host to most of Wales’ audiovisual history incuding BBC, ITV and S4C content. This project is in its final stages and has vastly improved access to this audiovisual material and also provided a springboard for developing access to other audiovisual content in the Library. The Screen and Sound team must now face new challenges in terms of technology, resources, rights, the material which is waiting to be collected, and new expectations from audiences. The team’s adoption of AI technologies features the need to support the Welsh language, and also the development of practical solutions with support from a collaborative project involving Kings University London, the BFI and other audiovisual archives in the UK.

The BBC’s archives hold decades of invaluable content but making it discoverable and reusable at scale is a significant challenge. This talk explores how AI-driven content enrichment can transform discovery and reuse workflows whilst keeping humans in the loop. We’ll examine approaches such as speech-to-text for indexing, face and object detection for tagging, and scene segmentation for improved navigation. Through real-world examples—like uncovering hidden interviews via transcripts and automating rights management with OCR—we’ll highlight how AI can be used to enhance accessibility, streamline workflows, and support editorial decision-making. Join us to explore how AI is being harnessed to support the BBC Library, unlocking new opportunities for discovery and reuse.

Technology is not only a tool—it’s a driver of transformation. In audiovisual information management, long based on human expertise, the arrival of algorithms marked a turning point. After an initial phase centered on automatic transcription and facial recognition, ATRESMEDIA’s Archive is now embracing generative AI.

This shift goes beyond innovation; it’s reshaping the identity of the audiovisual archivist. The next-gen professional is emerging as a hybrid: combining archival know-how with data skills, capable of supervising AI processes, ensuring metadata quality, and maximizing content value.

This evolution supports a larger goal: to redefine the archivist’s role as a key player in the content production ecosystem. At ATRESMEDIA, we strongly believe that the future of audiovisual archiving lies in professionals ready to collaborate with machines, shape new narratives, and actively contribute to the digital transformation of the media industry.

The Learning Curve: How the Groningen Archives leverage seminars like MMC to achieve great things in their relatively small (regional) archive. Why knowledge sharing is crucial and can be truly inspirational is illustrated through the experience of the Groningen Archives, where they have developed their own approach to incorporating humans into the loop of AI-generated metadata — with accuracy monitored by a dedicated group of volunteers. The Groningen Archives house the sound and video archives of both the local broadcaster (OOG TV) and the regional broadcaster (RTV Noord). These valuable collections form the foundation for the Ai metadata project called ‘The Sound of Groningen’.