DFL content production sees far-reaching AI-based upgrades
The DFL’s collaboration with Amazon Web Services (AWS) bears abundant fruit. Applications of Artificial Intelligence (AI) in particular create new options to tailor content to varying interests and make it available to Bundesliga fans around the world. This article describes two current examples: An automated localisation and translation solution, and an AI-based system to enrich video material with metadata that revolutionises content discoverability and usability.
Focus on fans: making content available in multiple languages
To provide international media partners with a variety of content formats suitable for their audiences, the DFL translates non-live content related to the Bundesliga and Bundesliga 2 into as many languages as possible. English, French, Portuguese and Arabic are most in demand.
However, the sheer volume of this media content poses a major challenge for translation. To overcome language barriers for most of its content and increase its reach, the DFL has developed an AI-based localisation and translation solution, supported by AWS. Labelled “Content OS” for the time being, the resulting application has passed the proof-of-concept test successfully.
It aims to enable media partners as well as internal DFL production teams to translate the text and sound of existing video clips into other languages automatically. Compared to the conventional translation process, this can reduce the required amount of time by 50 to 75 per cent, allowing the Bundesliga to create an offer that will allow more fans than ever around the world to experience the Bundesliga and Bundesliga 2 in their native languages.
Voice-to-voice in several steps
Being able to rely on a human product as the starting point for localisation was a key prerequisite for a successful prototype. “We do not want the result to be an ‘artificial’ product,” explains Luccas Roznowicz, Head of Strategic Cooperations at the DFL. “The original, whether sound track or subtitle, is always a human message. This puts the indispensable Human in the Loop right at the beginning. We were surprised how good the translation results were.”
We do not want the result to be an ‘artificial’ product
Luccas Roznowicz, Head of Strategic Cooperations DFL
Voice-to-voice translation – rendering spoken text in the original language as spoken text in the target language – is especially challenging. The processing software must first “understand” the original statement by matching it to a Large Language Model (LLM) while clarifying mumbled words or expressions. The resulting text is fed into a translation engine, whereupon the translated version is generated as spoken text.
“At first, the computer-generated voices sounded rather artificial,” reports Roznowicz. “But now the results are really good, if not perfect.” Perfectionism wouldn’t be helpful, he adds, since speed of delivery is of the essence: “If we are able to provide broadcasters with a product in their own language at a relatively good quality very fast, we help them keep their own effort to a minimum,” says Roznowicz.
The next step: a Minimum Viable Product
Meanwhile the solution has entered the next phase: developing a Minimum Viable Product, which undergoes further “training” to optimise the output step by step. The DFL is planning to let some initial test groups of partners and internal production teams order non-live content in specific target languages through the Media Portal. This localisation platform will become the common infrastructure for storytelling, offering a convenient combination of scalability, speed and quality.
Smart metadata generation
The DFL’s introduction of Intelligent Generation of Metadata (IGM) sets new standards in sport content processing and editing. As a native cloud-based system, IGM uses artificial intelligence to enrich video material with metadata automatically, taking content discoverability and usability to a new level.
“Our IGM addresses a key challenge facing all sport rights owners – generating the right metadata to realise the full potential of our unique content resources. AWS cloud technology and the use of artificial intelligence are the keys that unlock content by making it easy to find, search and use efficiently,” explains Luccas Roznowicz.
A new approach to processing and delivering content
IGM analyses video material using several AI models simultaneously to generate accurate, searchable metadata. The system can identify match events such as goals, emotions or unusual fan responses automatically. By combining conventional event identification with advanced semantic AI, IGM enables intuitive searches that can deliver answers to qualified queries such as “all of Harry Kane’s goals scored after the 80th minute” or “fans celebrating in the rain” within seconds.
The pathway to Smart Content
The DFL’s IGM sets new standards in content enhancement, demonstrating how AI can optimise the way football content is interpreted and produced.
IGM not only increases the efficiency of content production but also frees up human capabilities for other activities, such as creative storytelling. Equipped with a user-friendly graphical interface and search feature, IGM provides precise, AI-based metadata that let users find events such as goals, emotions or fan reactions. As a seamless solution, it revolutionises content generation by saving time and enabling media partners, clubs and editors to discover relevant content with ease. IGM has undergone the Proof of Concept (PoC) stage and will go live for the entire DFL media archives at a later time.
IGM harbours plenty of future potential
IGM can process any type of video material and determine the appropriate metadata for enrichment autonomously. It segments videos using AI, saving metadata for both, the individual segments and the entire video. Apart from the ability to identify events such as goals, the solution incorporates additional features including LLM-based text summaries, identification of humans and emotions, transcription, and vector embedding.