"AI commentary" is starting to mean too many different things at once. A voice clone reading translated English. A generated voiceover slapped onto a clip after the match has ended. A scoreboard turned into a spoken line. They are all being filed under the same label, and it is making the category harder to think about clearly.
We have been building Liat.ai for a while now with a fairly specific definition in mind. The kind of commentary we would want over a match we cared about, in the language we cared about, at the moment it actually mattered. Most of what is being marketed as AI commentary today does not meet that bar. This is a short note on what we think the bar should be, and why we have built Liat the way we have.
What Doesn't Count as AI Commentary
Three things tend to get called AI commentary that we do not think should be.
The first is translation with a voice clone. A real human team called the match in English, made every editorial decision, picked every emphasis, and an AI re-reads that script in Spanish or Portuguese with the original tone preserved. It is a useful product. But the football brain in that pipeline is still the original English booth. Every other language is downstream of one editorial voice that has already decided what mattered.
The second is highlight narration. A model writes a voiceover on a two minute clip after the match is over. Fine for socials, but the fans who needed commentary needed it live, during the chance, before the result was already public. By the time the highlight goes out, it is a recap.
The third is the template reader. A line generated from an event stream that says "Goal scored by Player X in the Y minute." A spreadsheet could read that out. There is no understanding of the game behind the voice, and you can hear it.
What Real AI Commentary Has to Do
For us, a system has to do four things before it can call itself an AI commentary service.
It has to see the match. Not consume an event feed, actually see it. Our system runs computer vision over the live video, tracking the ball, the players, jersey numbers, field positions, contests, dribbles, carries, the lot. That is what gives the voice something to actually say.
It has to understand the match. A 3-1 in stoppage time of a relegation playoff is not the same event as a 3-1 in a friendly. Our commentary layer pulls in league standings, head-to-head history, player form, club rivalries, and the weight of the specific minute the chance arrives in. Without that, an AI is just narrating geometry.
"Anything that does not do all four is not AI commentary. It is automation with a voice on top."
It has to say it fast enough to matter. End to end, under a second from event on the pitch to spoken word. If the call lands after the broadcast graphics catch up, it is a recap dressed as commentary.
And it has to sound like it belongs there. Authentic voices, two of them working together when the match calls for it, generated natively in each language with the rhythm and idioms of that language. Not English thrown through a translation layer.
Built for the Matches No One Calls
Most matches in the world have no commentary on them.
That is the part that drives almost every design decision we make. Sunday morning grassroots in Lagos. A women's reserve game in Madrid. A youth tournament in Ho Chi Minh City. A second division match in Lima that is broadcast nowhere. The reason these games go untold is not that no one cares. It is that the economics of human commentary make it impossible to reach them.
A live human team for a single match runs above a thousand dollars in talent, production, and infrastructure. Liat costs under ten. That is two orders of magnitude, which is a different category of product. It is what lets us cover the matches that have always been silent, in the languages the fans actually speak.
Customisable Down to the Match
The other reason for being precise about what AI commentary is, is that the product has to flex with the match.
A penalty in a school final does not sound like a penalty in a cup final, and we should not pretend otherwise. So everything about the voice is configurable. The language. The accent. Two voices working together, or a single voice working alone. The level of tactical depth. The crowd bed. Sponsorship constraints. The house style of a broadcaster. A focus on a specific player or a specific team. Intensity dialed up for a derby and dialed down for a friendly.
A real AI commentary service cannot ship one voice and one style and call it done. It has to fit the match it is calling.
A fan in São Paulo turning on a second division match does not care that there is an AI voice on top of a different match somewhere else. They care that the match in front of them is being told, in their language, while it is still happening.
That is the service we set out to build. Liat watches the match itself, understands the moment it is in, and speaks in the language the fans speak, for any match on the planet. That is what we mean by the first true AI sports commentary service.
You can hear it for yourself on the demo page, read more about how the system works on the technology page, or get in touch about coverage.