ElevenLabs Introduces Scribe v3 and Scribe v3 Live for Speech Transcription
ElevenLabs announced new models for transcribing recordings and live conversations. The company reports improved speaker identification and lower error rates compared with the previous version, but the results come from its internal testing.
ElevenLabs announced ElevenLabs Scribe v3, a new model for transcribing recorded audio, and Scribe v3 Live for real-time speech transcription on November 10, 2026. The new products expand the company’s offerings for applications such as meeting notes, voice agents, and contact centers.
The models have the identifiers scribe_v3 and scribe_v3_live. Scribe v3 is available through the web interface and API, while the Live version is intended for the realtime API and live conversations.
What ElevenLabs Scribe v3 Brings
According to ElevenLabs documentation, Scribe v3 adds better labeling of individual speakers, number formatting, and the option to select secondary languages. These features are particularly important for recordings involving multiple people or where incorrectly transcribed numbers and names could affect further automated processing.
Scribe v3 Live is a separate model focused on continuous transcription. The company describes it as a tool for live conversations—scenarios in which an application needs to work with spoken input without waiting for the entire recording to finish.
For existing integrations, switching in the API may be possible by changing the model identifier. However, the announcement does not specify whether developers could encounter differences in model behavior or output during migration.
Company Testing Reports Lower Error Rates
ElevenLabs states that Scribe v3 achieved 27% fewer word errors than Scribe v2. This result comes from the company’s internal evaluation using a modified version of the FLEURS benchmark in 102 languages.
It is therefore a vendor-reported figure, not an independently verified comparison across the market. The available information does not show that the new model is more accurate than all competing systems. The claimed improvement on real recordings, including speaker identification, has not yet been externally verified.
For speech transcription, average error rates alone do not capture every practical situation. Microphone quality, background noise, accent, language switching, and the number of speakers may all matter. For live services, latency is also important, but the fact sheet does not provide comparative results.
Uncertainty Around ElevenAgents
The ElevenLabs blog states that Scribe v3 Live is available in the ElevenAgents service. At the same time, the current support page identifies Scribe v2 Realtime as the default model in ElevenAgents. Based on the available documentation, it is therefore unclear what stage deployment of the new realtime version has reached in this service.
ElevenLabs Scribe v3 thus brings updated models for batch and live transcription, with the company reporting improved processing of speakers, numbers, and multilingual content. The accuracy reported in internal testing will need to be compared with independent tests.
Another point to watch is the release of the promised real-recording dataset and full comparison methodology. Tests of accuracy, latency, and diarization across different languages will also be relevant, as will confirmation of when Scribe v3 Live becomes the default model in ElevenAgents.
Sources
- ElevenLabs Product Blog – Confirms the announcement, release date, intended use of both models, and the company’s internal measurement results.
- ElevenLabs Documentation – What is Speech to Text? – Confirms the model identifiers, availability through the web and API, and the stated new Scribe v3 features.
Verified and updated: 11/11/2026 06:32
