Live from Paris 2024: How OBS is using AI to speed up the creation of highlights

Automated highlights generation: OBS tech on display in the IBC, Paris

Olympic Broadcasting Services (OBS) has embraced AI for its coverage of the Paris 2024 games, deploying tools to speed up the content creation process including an automated highlights platform and an Al-driven tool to analyse live commentary transcripts to create summaries for editors.

Speaking with SVG at the IBC in Paris during the Games – when OBS was on track for delivering an unprecedented 11,000 hours of content – OBS director of broadcast engineering Guillermo Navarro explained how the host broadcaster had developed a platform which uses AI technology to identify and compile key moments into reels tailored for multiple platforms.

Since 2018, OBS has tried different ways of tagging athletes to assist with editing, but with facial recognition not permitted in France (AI video surveillance is only allowed as a trial during the Paris 2024 by security services) OBS tried other methods such as OCR (optical character recognition) on athletes’ jerseys or bibs.

“That comes with its own complexity, because you need to isolate the person from any other people within the shot,” Navarro said. “Plus, the bib might not be perfectly aligned, so instead of ‘16’, you might see ‘6’. So it was causing lots of false positives.

“In the end, our main goal was to assist with the generation of highlights in an automated way, and this is how in the end, and for the first time, we came up with a tool that understands each of the sports. Together with a data feed, it provides a selection of cuts according to the rules that the user or the platform predefines.”

The highlights could be related to particular categories, rather than traditional highlights that summarise a match or event, and could instead be related to a game, specific team or player or even particular actions, explained Navarro.

OBS initially started with 14 sports, but the interest of broadcasters encouraged OBS to increase that number to 31, including major sports like athletics, gymnastics plus the likes of badminton and wrestling. “So it’s a big variety of sports in terms of field of play and the type of competition,” said Navarro.

OBS has teamed up with official AI platform partner Intel on the project, which uses LLMs (large language models) – the same technology that is behind ChatGPT and OpenAI. For each sport, it requires a new model and training – with the time taken to train the model dependent on the complexity of the sport and the data that accompanies footage of the sport. Athletics, for example, is complex because of the many components and disciplines, whereas a sport like hockey is relatively straightforward.

Looking ahead to the Milano Cortina 2026, Navarro said OBS is looking “tentatively” at using the platform to cover all sports at the Winter Games.

OBS has also developed an Al-driven tool to streamline the editing process, with Al used to analyse live commentary transcripts to create summaries – also using LLMs – which are then used to generate edit markers. For two hours of content, it might generate 50 or 60 summaries, and then a final summary to provide an overview of all of the action.

“It’s similar to if you had your friend, or your colleague, just watch the match, and you ask them, ‘Okay, can you tell me what just happened? What were the highlights and what was relevant?’” explained Navarro.

The summaries are linked to timestamps, and via a plugin for Adobe Premiere they are dumped on to an editing timeline to create an initial rough cut for a producer.

“The editor and the producer will say, ‘okay, trim here, take out this,’ and maybe they can be done in 20 or 30, minutes, whereas before, a producer would need to watch the entire competition which might last two or three hours, take notes, and then sit with an editor,” said Navarro.

He added that it was difficult to quantify the amount of time saved, or the uplift in content produced, because it was a relatively late – in Olympics planning terms – addition to the production process.

“We don’t have figures yet, because we have our full regular team in place,” said Navarro. “But this allows us to start putting the building blocks for something in the near future, when we can define a new scope for the production team. The idea is not to start cutting people or things like that; there are so many opportunities for new content that go untapped because we don’t have more resources.”