What ElevenLabs released
On 10 November 2026, ElevenLabs published a post by Min Kim.
The page also lists a last update on 13 November 2026.
The post launches two speech-to-text models.
Scribe v3 is for audio that is already recorded.
Scribe v3 Live is for audio that is still coming in.
The post says both are the default in the ElevenLabs app.
Scribe v3 Live is also available in ElevenAgents.
Both models are in the ElevenLabs API.
An existing project changes the model id to scribe_v3 or scribe_v3_live.
The post names no price and no plan.
What the 27 percent measures
ElevenLabs says Scribe v3 makes 27 percent fewer word errors than Scribe v2.
The comparison covers 102 languages.
It uses an internal version of a public test called FLEURS.
FLEURS uses short, clean readings.
Each clip has one speaker.
The post says v3 improved in every language in that test.
It says the word error rate fell by 42 percent for German.
It fell by 35 percent for Spanish and by 23 percent for English.
It fell by 34 percent for Swahili, 29 percent for Tamil, and 28 percent for Thai.
The post does not name Filipino, Tagalog, Cebuano, or Taglish.
ElevenLabs also says v3 made fewer word errors than the other recorded-audio models it tested.
Those comparisons use only languages both models support.
ElevenLabs says it scored each language first, then averaged them.
Each language counts the same in that average.
A score that pools every clip into one number is a different test.
Speaker labels and live audio
The post says Scribe v3 has the most accurate speaker labels of the models it tested.
On recordings with three or more speakers, it says label errors fell by 17 percent from v2.
Across its tests, it says those errors fell by 12 percent from v2.
Scribe v3 Live adds speaker labels to real-time audio for the first time.
The post says Live makes 39 percent fewer label errors than the next real-time model it tested.
It also describes a set of real meetings, calls, and voice notes in 16 languages.
On the English part of that set, it says the Live label error is 3.1 percent.
The other real-time models in that test scored 5.9 to 7.3 percent.
The post says a standard label score also counts missed words.
A second score lines the words up first.
The post says that second score is not comparable with the batch results.
For word errors, the post separates two scores.
One score counts an error only when the meaning changes.
ElevenLabs says that score is 43 percent lower than Scribe v2 Realtime.
It says the ordinary word error rate is 31 percent lower.
The post says the Live error rate was lower than every other real-time model it tested.
Terms you can supply
Scribe v3 accepts up to 1,000 terms you expect to hear.
The examples in the post are names, acronyms, and jargon.
You can ask for numbers as digits or as words.
You can also give a formatting rule in ordinary English.
The post’s example is to write every date as a year, then a month, then a day.
Scribe v3 Live adds a mode for talk that switches language mid-sentence.
It also adds a larger term library for words that might come up.
The post says that library holds thousands of terms.
The model checks the library only when it hears a likely match.
Where the audio runs
The post says both models run in the United States by default.
Enterprise customers can use separate environments in the European Union, India, and Singapore.
A setting called Zero Retention Mode deletes the audio and the text when a request finishes.
The post also says a medical version of Scribe v3 Live is in ElevenAgents.
It does not say that version is offered in the Philippines.
It names no hospital and no fee.
Why it matters in the Philippines
A podcast, a news interview, or a support call starts as audio.
A wrong name or a wrong number is easy for a summary to repeat.
ElevenLabs gives a scale in the same post.
A one-hour meeting at 150 words a minute is about 9,000 words.
The gap between a 4 percent and a 2 percent word error rate is 180 wrong words.
That example is not a measured result for Scribe v3.
It shows why a smaller error rate can matter on a long recording.
For a Filipino show or a contact center, the open question is the language.
The 102-language test does not list a Philippine language.
A team should also read where the audio is processed before it uploads a customer’s voice.
What to do next
- Open the 10 November post and read the test note before you repeat the 27 percent.
- Check your own ElevenLabs account for Scribe v3, because the post does not name a plan.
- If your audio is Filipino or Taglish, transcribe one short clip and read the names yourself.
- If the recording belongs to a client, read the United States default before you upload it.
- For an API project, the post says to set the model id to scribe_v3 or scribe_v3_live.
How this was checked
The dates, the model names, the error figures, and the 1,000-term limit come from the ElevenLabs post.
The app default and the United States default come from that same post.
The post was published on 10 November 2026.
It lists an update on 13 November 2026.
This desk opened the post on 14 November 2026.
The notes for Filipino creators and contact centers are this article’s reading.
ElevenLabs did not publish those steps.
The app was not opened for this check.
Put this into practice
The accuracy figures come from ElevenLabs. These guides show how to check a company test, and how to compare this tool with the one you use now.
Sources & further reading
Source note. The linked sources support the news in this article. The explanations and suggested next steps are from Pinoy AI Works.


