ElevenLabs introduced two new speech-to-text models on November 10, 2026. Scribe v3 transcribes recordings, while Scribe v3 Live works in real time. Both now run as the default model in the ElevenLabs app.
For German, the manufacturer puts the improvement over Scribe v2 at 42 percent fewer word errors. It was measured on an internal version of the FLEURS benchmark, which tests short, clean read speech. This test does not cover noisy interviews.
What does Scribe v3 improve over Scribe v2?
Across all 102 languages, Scribe v3 delivers 27 percent fewer word errors than its predecessor, according to ElevenLabs. Speaker attribution (diarization), meaning the question of who said what, shows 12 percent fewer errors, and even 17 percent fewer for recordings with three or more voices. Also new are keyterms: you can give the model up to 1,000 terms in advance, such as brand names, abbreviations, or technical terms. You write formatting rules in simple English sentences, for example "write every date as YYYY-MM-DD", and the model outputs numbers either as digits or as words.
Scribe v3 in numbers
Figures from ElevenLabs compared to Scribe v2, introduced on November 10, 2026
- 42 %fewer word errors in German
- 27 %fewer word errors across 102 languages
- 17 %fewer errors in speaker attribution with three or more voices
- 1,000technical terms you provide in advance
The value for German comes from short, clean read speech, not from noisy recordings.
Source: ElevenLabs
According to the ElevenLabs post, Scribe v3 Live delivers 43 percent fewer errors that change the meaning (semantic errors) than Scribe v2 Realtime. For the first time, the model also recognizes in real time who is currently speaking. According to ElevenLabs, speaker mix-ups are 39 percent lower than those of the next-best real-time model in the test. The model recognizes whether a speaker has finished from the conversation context and not from silence alone.
What does Scribe v3 cost, and where is the data processed?
Via the API, ElevenLabs bills per hour of audio, according to the pricing page. Scribe v3 Live costs ten percent more than Scribe v3. Keyterms and formatting rules that edit the text afterwards are charged by the provider additionally. The page lists the amounts in US dollars; there is no euro price. You switch existing integrations over with the model ID scribe_v3 or scribe_v3_live.
By default, both models run in the US. ElevenLabs offers separate environments in the EU, India, and Singapore to companies with data residency requirements. Zero Retention Mode, a mode without storage, deletes audio and text as soon as a request is completed. Whether individuals can use the EU environment is not answered by the post. It does not state prices for the app plans either.
What this means for you
If you transcribe interviews, podcasts, or subtitles, it is best to try Scribe v3 with a German recording on which your previous tool failed. Enter product names and proper names in advance as keyterms and compare both versions word for word. The 42 percent comes from clean read speech; your recordings with reverb or background noise may turn out differently.
If you process recordings from clients, clarify before use whether the assignment requires processing in the EU. The default is the US, and ElevenLabs mentions the EU environment only for enterprise customers.
Calculate in advance with your hours of material, because keyterms and formatting rules increase the hourly price. Our assessment: a comparison test with your own material is worthwhile before you commit to a model.
ElevenLabs introduced Scribe v3 and Scribe v3 Live on November 10, 2026.
Sources
