ElevenLabs launches Scribe v3: transcription with 27% fewer errors
ElevenLabs launched Scribe v3 and Scribe v3 Live on Tuesday, its new speech-to-text models. Here is what the numbers say, what they leave out about Portuguese, and what changes for anyone captioning clips.
ElevenLabs launches Scribe v3: transcription with 27% fewer errors
On Tuesday (November 10), ElevenLabs launched Scribe v3 and Scribe v3 Live, its new speech-to-text models: the first for recorded audio, the second for real time. The promise is 27% fewer wrong words than Scribe v2 on average across 102 languages, and for anyone captioning clips the most interesting part is not even that: it is speaker separation, which improved precisely in conversations with three or more people.
Both models are already in the API and have become the default in the ElevenLabs app. The announcement has no country restrictions, so it applies from today to people in Brazil too.
What did ElevenLabs announce?
Two models, with gains measured against the company's own previous generation. All the numbers below come from the launch post and the models documentation, not from independent testing.
| Scribe v3 (recorded audio) | Scribe v3 Live (real time) | |
|---|---|---|
| API id | scribe_v3 |
scribe_v3_live |
| Word error vs. v2 | 27% fewer, across 102 languages | 31% fewer (43% fewer meaning errors) |
| Speaker separation | 12% fewer errors; 17% fewer with three or more people | First time in real time |
| Keyterms | Up to 1,000 | 50 per connection, plus a library for large vocabularies |
| Audio in more than one language | Accepts secondary languages set in the call | Multilingual mode, switching mid-sentence |
On the FLEURS test set average, the word error rate dropped from 16.94% to 12.38%, according to the documentation. Among the highlighted languages, German improved 42%, Spanish 35% and English 23%. The company says v3 improved in every language tested and makes fewer errors than every competitor it compared, but the announcement text only says "other popular models", without naming them.
One detail for those who read with a magnifying glass: the post mentions 43% fewer meaning errors in Live, while the documentation says 44%, measured across 73 languages. The order of magnitude is the same, but it is the kind of difference that shows these numbers are in-house.
What about Portuguese?
The announcement does not mention Portuguese. The good news is that it was already starting from a high place: in the speech-to-text documentation, Portuguese appears in the "Excellent" tier for Scribe v2, with up to 5% word error.
That changes how to read the 27%. In a language that already made few mistakes, there is less room to improve on average, and the gain tends to show up where the old model stumbled: proper names, slang, people talking over each other. That is what matters in a clip caption, and FLEURS, which is read speech, does not measure it. I wrote about this in how to test auto-caption accuracy: the only number that counts is the one from your own audio.
What changes for people who clip podcasts and livestreams?
The gain that shows up most on screen is speaker separation. In a podcast with three guests, the classic error is the caption attributing a line to the wrong person, and then a clip that alternates color or position by speaker falls apart.
A 17% drop in that error in recordings with three or more voices is what matters most to anyone building vertical framing for a podcast with two or three people. v3 supports up to 32 speakers per file, according to the documentation.
In real time, Live brings two new things. One is knowing who is speaking while the person is still speaking: in English, ElevenLabs measures 3.1% error in speaker separation, against 5.9% to 7.3% for the other models it tested. The other is end-of-turn detection that uses the context of the conversation, not just silence, which should cut fewer sentences in half. For anyone thinking about clipping livestreams in real time, this kind of improvement reduces the rework of adjusting where a clip starts.
Keyterms are the underrated feature. You pass the list of names the model should recognize (the streamer's nickname, the game's name, the sponsor's brand) and it stops inventing spellings. v3 fits up to 1,000 terms, according to the documentation. For a single creator's clips channel, a fixed list of 30 names already solves a good share of the errors you fix by hand today.
What we still do not know
Three things were left out of the announcement:
- Price. As of this post, the API pricing page still listed only Scribe v2 ($0.20 per hour of audio) and v2 Realtime ($0.22), with keyterms and transcript editing billed separately.
- Live latency. v2 Realtime had a median of about 150 milliseconds; v3 Live has no published number.
- Results per language outside the highlights. Without the Portuguese number, the promise of "improved in all of them" is what we have.
What to do now
If you use the ElevenLabs API, the switch is only the model id, according to the company. Before flipping the key in production, run the same minute of audio through v2 and v3 and compare word by word, counting errors in proper names and numbers separately. It is half an hour of work and it answers the question the announcement does not.
If you do not integrate any API and just want captions on your clips, what counts is the tool you use. On Cut.Pro, the AI cuts the video, generates the captions and you fix them word by word in the editor before publishing. The one-minute test works the same for us and for any competitor. If the captions you post today already come out right, nothing changes. If you still lose time fixing guests' names, well-made captions hold retention, and it is worth demanding that improvement from your tool.
