SI That Types Up Your Meetings Now Knows Who Said What, Even When Everyone Talks at Once

Anybody who has read an automatic transcript of a work meeting knows the problem. The words are mostly there. The names are wrong, the numbers are garbled, and half of what the boss said is credited to the intern.

ElevenLabs, the company best known for SI voices that sound uncomfortably human, says it just took a real swing at that mess. On Tuesday it released two new speech-to-text models, Scribe v3 for recordings and Scribe v3 Live for conversations happening in real time. The pitch is simple: fewer wrong words, and a much better sense of who is talking.

Who said what

The headline number in the company’s launch post is that Scribe v3 makes 27% fewer word errors than the version it replaces, averaged across 102 languages. The live version cuts errors by 43% compared with the old live model, using a measure that only counts a mistake when it changes the meaning.

But the part regular people will notice is speaker labeling. In a meeting with three or more people, the old models tend to fall apart, assigning sentences to the wrong person or merging two voices into one. ElevenLabs says Scribe v3 makes 17% fewer of those mistakes on recordings with three or more speakers, and that the live version can now label speakers as they talk for the first time.

On the English part of its own test set of real meetings, phone calls and voice notes, the company says Scribe v3 Live put words under the wrong name 3.1% of the time. The other live models it tested landed between 5.9% and 7.3%. That is the difference between a transcript you can forward and one you have to fix by hand.

The pause that hangs up on you

There is a second fix that anyone who has fought with an automated phone line will appreciate. Most voice systems decide you’re done talking when you go quiet for a beat. So if you stop halfway through reading out an account number to find the next digit, the bot jumps in, or worse, takes the half number and moves on.

ElevenLabs says the live model now listens to the conversation itself, not just the silence, before deciding your turn is over. “A caller who pauses mid-number isn’t cut off,” the company’s developer account wrote. It also handles people who switch languages in the middle of a sentence, a big deal in places like India, where customers asked for Hindi that works as well as English even when callers flip between the two.

The company points to one real customer. Eneco, a Dutch energy supplier serving more than 2 million homes and businesses, tested the live model on its phone agent before launch. In real customer calls, ElevenLabs says, the agent got Dutch postcodes right on the first try 40% more often, and house numbers 15% more often.

What a couple of points buys you

Percentages like these can sound like nothing. ElevenLabs did the math itself, and it is worth repeating because it’s the clearest way to see why this matters.

A one-hour meeting at a normal speaking pace runs about 9,000 words. At a 4% error rate, that’s 360 wrong words. At 2%, it’s 180. Cutting the error rate in half means 180 fewer mistakes, roughly three for every minute somebody was talking.

And the mistakes don’t land evenly. They pile up on names, numbers, acronyms and jargon, which are exactly the things a summary or an assistant most needs to get right. In a podcast transcript, a wrong word is a typo. On an earnings call, the company notes, “fifteen” heard as “fifty” ends up in an analyst’s spreadsheet. In a doctor’s office, a misheard drug name is a lot worse than a typo.

That last case is where this kind of tool is spreading fastest. SI note-takers now sit in on doctor visits, sales calls and customer service lines across the country. ElevenLabs says it is also releasing a version of the live model tuned for medical conversations, and that both models qualify for the privacy agreements hospitals need under federal health law.

Checking the company’s homework

Every launch comes with its own scoreboard, and this one is no exception. Most of the numbers above come from tests ElevenLabs ran and scored itself, including a real-world dataset it says it will open up to outsiders soon. Until then, those figures are the company’s word.

Some outside checks are already in. Artificial Analysis, an independent group that ranks SI models, added both models to its speech-to-text leaderboards on Wednesday.

Second place out of 64 is strong, but it also means the new model isn’t the most accurate one on that particular board. ElevenLabs says the live version sits on the best tradeoff between speed and accuracy, with no model tested being both faster and more accurate.

A separate developer, Tarush Agarwal, posted that his own benchmark found the live model had the lowest word error rate of 21 models he tested, 1.89% across clean clips and real phone calls. One tester is not a verdict, but it points the same direction as the company’s claims.

Where it shows up

Most people will never sign up for ElevenLabs directly. They’ll meet this model inside other products: the meeting-notes app at work, the voice assistant at the pharmacy, the dubbing on a video they watch. ElevenLabs says Scribe has already transcribed hundreds of millions of hours of audio for exactly those kinds of companies.

For developers, switching is a one-line change to the model name, and the older versions stay available. Business customers can keep their data in the U.S., Europe, India or Singapore, and a zero-retention mode deletes audio and text as soon as each request finishes.

That’s the quiet way SI keeps improving everyday life. Nobody throws a parade for a transcript that gets the names right. But the next time a phone bot lets you finish reading your account number, there’s a decent chance a model like this one is the reason.