Share𝕏inWA

ElevenLabs Scribe v3: Sharper Transcripts at Same Price

TechorCoffee
TechorCoffee
Contributing Writer

Published (Updated )

ElevenLabs Scribe v3: Sharper Transcripts at Same Price

ElevenLabs Scribe v3: Sharper Transcripts at Same Price

Transcription is one of those jobs where a small improvement matters a lot. A few wrong words in a meeting note can change what was agreed. A missed speaker label can make a transcript hard to follow. A name spelled the wrong way, again and again, can make the whole document feel unreliable.

ElevenLabs released Scribe v3 and Scribe v3 Live on 10 November 2026. The blog post announcing them is dated 10 November. The developer changelog lists the release under 9 November. That is a one-day publication-date difference between two ElevenLabs pages, not a separate event. It is worth noting, because dates get copied around quickly, and a second date can look like a second launch when it is not.

The two models do different jobs. Scribe v3 is for recorded audio. Scribe v3 Live is for low-latency real-time transcription. Both are now the default speech-to-text models in the ElevenLabs app and API. Their model IDs are scribe_v3 and scribe_v3_live.

That default status matters. Existing users do not need to hunt for a new setting to get the new models. New users get them from the start. The change is in the model behind the button, not in a new product to learn.

Image 1 — after the introduction, before "Fewer word errors, on ElevenLabs' own tests"

ElevenLabs Speech to Text product page, showing Scribe v2, the model Scribe v3 replaces

Fewer word errors, on ElevenLabs' own tests

The headline claim is about accuracy. ElevenLabs says Scribe v3 makes 27% fewer word errors than Scribe v2 across 102 languages, measured on its own version of FLEURS. The word error rate, it says, fell from 16.94% to 12.38%.

Those numbers need careful reading. They are ElevenLabs' own benchmark results, with no independent verification. The test is the company's own version of FLEURS, not a public test run by someone else. That does not make the numbers false. It does mean they come from the vendor, testing its own model, on a test it runs itself.

A 27% drop in word errors is a large claim. In plain terms, it means roughly a quarter fewer wrong words than the previous model, on that test, across those 102 languages. A word error rate of 12.38% still means about one word in eight is wrong, on average, in that test. That is better than 16.94%, and it is still far from perfect. Transcripts will still need checking, especially where the words matter.

Scribe v3 Live, the real-time model, is claimed to make 43% fewer errors than Scribe v2 Realtime. Again, this is ElevenLabs' own figure. Real-time transcription is harder than recorded transcription. The model has to keep up as the words arrive, with no chance to go back and re-listen to a whole sentence. A large improvement there, if it holds up in real use, would matter for live calls, voice agents, and captions.

The honest way to use these figures is as a guide to what ElevenLabs expects, not as a promise about your audio. Your recordings will differ from a benchmark set. They will have different speakers, different accents, different background noise, and different subject matter. The only test that settles how Scribe v3 does on your audio is your audio.

Who said what, even in real time

Speaker labels, known as diarization, are the second big change. Scribe v3 has 12% fewer diarization errors overall than Scribe v2, according to ElevenLabs, and 17% fewer on recordings with three or more speakers.

That second figure is the more useful one. Two-speaker recordings are the easy case. Three or more speakers, talking over each other, interrupting, or speaking in short turns, are where transcripts usually fall apart. A 17% drop in errors there would make a real difference to a meeting transcript, a group interview, or a panel discussion.

Scribe v3 Live adds diarization to real-time transcription for the first time. Until now, live transcription could give you the words as they came, but not a reliable label for who said them. Now, ElevenLabs says, the live model can do both at once. For a live meeting, a support call, or a voice agent, that changes what the transcript is good for. You can follow the conversation as it happens, not just read it back later and try to work out who was speaking.

A version of Scribe v3 Live tuned for medical conversations is available in ElevenAgents. That is a separate point, and it is worth keeping separate from the general claims. Medical talk has its own vocabulary, and a model tuned for it is aimed at a specific use, not at every recording.

Image 2 — after "Who said what, even in real time", before "Keyterms: tell it the hard words first"

ElevenLabs speech-to-text interface showing a transcription with a speaker label and timeline

Keyterms: tell it the hard words first

One of the most practical new features is keyterms. Users can supply up to 1,000 terms in advance. Those can be brand names, abbreviations, or technical terms.

Anyone who has transcribed a meeting knows why this matters. A product name, a company name, or a piece of jargon is exactly the kind of word a general model gets wrong. It hears something close, and writes something plausible, and the error repeats every time the word comes up. Giving the model the right spelling in advance is a simple fix, and it targets the errors that are most annoying in practice.

A thousand terms is a generous limit. It is enough for a company's product list, a project's acronyms, and the names of the people in the room. It does not make the model perfect on everything else, but it gives users a direct way to fix the words they care about most.

The price did not change

Prices did not change. Scribe v3 costs $0.20 per hour of audio through the API, and Scribe v3 Live costs $0.22 per hour. Those are the prices listed on ElevenLabs' API pricing page.

Holding the price is a deliberate choice, and it is part of the story here. A new model often comes with a new price. Here, the company is saying the improvement comes at the same rate as before. For anyone already paying for Scribe v2, that removes one reason to hold back. The cost per hour stays the same, while the claimed accuracy goes up.

That does not mean transcription is free, or cheap for very large volumes. At $0.20 an hour, a hundred hours is $20. A thousand hours is $200. The arithmetic is simple, and it is worth doing before a large job. The point is only that the rate itself has not moved.

Where the audio is processed

Both models run in the US by default. Enterprise customers can use isolated environments in the EU, India, and Singapore.

That detail will matter most to teams with rules about where data goes. A default US location is fine for many users, and a problem for others. The isolated environments give larger customers another option, in three regions, without changing the model itself.

It is a short point, but it belongs in any honest summary. Accuracy and price are not the only things people check before they send audio to a service. Where that audio runs, and under what arrangement, is part of the decision too.

Image 3 — after "Where the audio is processed", before the comparison table

Illustration for ElevenLabs Scribe, showing a robot speaking in several languages

ModelForHeadline claim from ElevenLabsAPI price per hour
Scribe v3Recorded audio27% fewer word errors than Scribe v2$0.20
Scribe v3 LiveReal-time audio43% fewer errors than Scribe v2 Realtime$0.22

An outside test, on real phone calls

There is one independent test to mention, and it is limited. Voice Arena ran Scribe v3 Live on more than 70 hours of real phone calls, and ranked it first in Romanian and Indian English, and third in US English.

That is a useful check, because phone calls are messy in a way benchmarks are not. People talk over each other, lines crackle, and accents vary. A test on real calls is closer to real use than a test on clean, read speech.

It is also a narrow check. It covers Scribe v3 Live, not Scribe v3. It covers a handful of languages, not all 102. And a ranking of first or third tells you where the model stood in that test, not how it will do on your calls. Still, it points in the same direction as ElevenLabs' own figures: better than the previous model, and strong in some languages.

A day later, ElevenLabs and NVIDIA said NVIDIA's Nemotron 3.5 Super VL model is available in ElevenAgents. That is a separate announcement, about a different model in a different product, and it should not be read as part of the Scribe v3 launch. It is worth a line here because it came a day later, and because it shows ElevenAgents, where the medical version of Scribe v3 Live sits, is also where other new models are landing.

The bottom line is simple. Scribe v3 and Scribe v3 Live claim to be sharper than the models they replace, with better speaker labels, a useful keyterms feature, and the same price. All the accuracy figures are ElevenLabs' own benchmark results, with no independent verification. Try them on your own audio, check the transcripts where it matters, and judge them on that, not on the launch numbers alone.

FAQs

What is the difference between Scribe v3 and Scribe v3 Live?

Scribe v3 is for recorded audio. Scribe v3 Live is for low-latency real-time transcription, and it adds diarization to real-time transcription for the first time.

How much more accurate is Scribe v3 than Scribe v2?

ElevenLabs says Scribe v3 makes 27% fewer word errors than Scribe v2 across 102 languages, with the word error rate falling from 16.94% to 12.38% on its own version of FLEURS. These are ElevenLabs' own benchmark results, with no independent verification.

What are the model IDs for the new models?

The model IDs are scribe_v3 for Scribe v3 and scribe_v3_live for Scribe v3 Live. Both are now the default speech-to-text models in the ElevenLabs app and API.

How much does Scribe v3 cost?

Prices did not change. Scribe v3 costs $0.20 per hour of audio via the API, and Scribe v3 Live costs $0.22 per hour.

What are keyterms?

Keyterms let users supply up to 1,000 terms in advance, such as brand names, abbreviations, or technical terms, so the model is more likely to get those words right.

Where do Scribe v3 and Scribe v3 Live run?

Both models run in the US by default. Enterprise customers can use isolated environments in the EU, India, and Singapore.

Was Scribe v3 released on 9 or 10 November 2026?

The blog post is dated 10 November 2026, and the developer changelog lists the release under 9 November 2026. That is a one-day publication-date difference, not a separate event.

Was this article helpful?

TechorCoffee

Written By

TechorCoffee

Contributing Writer

TechorCoffee is a technology reporter and software creator writing reviews and tutorials for TechOrCoffee.

Comments (0)

Sort by:

Leave a Comment

Loading comments...

Read Next