ElevenLabs released two speech-to-text models on November 10: Scribe v3 for recorded audio and Scribe v3 Live for real-time audio. Both are now the defaults in the ElevenLabs app. In the API they are scribe_v3 and scribe_v3_live, and an existing Scribe v2 integration only needs the new model ID. Scribe v3 Live also powers voice agents built with ElevenAgents.
Key changes
- Fewer errors: ElevenLabs says Scribe v3 makes 27% fewer word errors than v2 across 102 languages on the FLEURS benchmark, with the largest drops in German (42%) and Spanish (35%) and 23% in English.
- Real-time accuracy: Scribe v3 Live makes 31% fewer word errors and 43% fewer meaning-changing errors than Scribe v2 Realtime.
- Speaker labels in real time: Live can now tell speakers apart as it transcribes. In ElevenLabs’ English test its error rate was 3.1%, against 5.9% to 7.3% for the other real-time models it tested. Batch speaker labelling improved 12%, and 17% with three or more speakers.
- Vocabulary control: keyterm prompting takes up to 1,000 terms, and Live adds a keyterm library of thousands of terms that is checked only when a word looks like a match.
- Conversation features in Live: turn detection that uses the conversation, not only silence; language hints and a mode for speakers who switch language mid-sentence; numbers written as digits or words.
Data and compliance
Audio is processed in the US by default, with isolated environments in the EU, India and Singapore. Zero Retention Mode deletes audio and text when a request finishes. ElevenLabs lists SOC 2, ISO 27001 and PCI DSS Level 1, and HIPAA eligibility with a business associate agreement. A version of Scribe v3 Live tuned for medical conversations is planned for ElevenAgents.
What it means for users
For meeting notes, interviews and subtitles, the main change is fewer corrections, especially in non-English audio. For call centres and voice agents, real-time speaker labels and better handling of names, postcodes and product terms matter more: an energy company quoted by ElevenLabs says postcodes were recognised on the first try 40% more often. All accuracy figures come from ElevenLabs, and the announcement does not give prices, so check your plan before switching high volumes. A simple test is to run the same ten recordings through v2 and v3, including your hardest accents and product names, and compare how many edits each transcript needs.
Related tools
See ElevenLabs on ToolAI, and compare it with AssemblyAI, Otter.ai, Fireflies.ai and Descript.