ElevenLabs Improves AI Speech Recognition Accuracy... Cuts Errors by Up to 43%
- Input
- 2026-11-19 08:16:18
- Updated
- 2026-11-19 08:16:18

According to industry sources on the 19th, ElevenLabs unveiled Scribe v3, which converts recorded speech into text, and Scribe v3 Live, a real-time speech recognition model. Both models are available through ElevenAPI for developers, while the real-time model is also offered through ElevenAgents, the company’s enterprise voice and chat agent platform.
ElevenLabs said its own evaluation found that Scribe v3 reduced word recognition errors by 27% compared with its predecessor, Scribe v2. Scribe v3 Live, which is specialized for real-time speech recognition, cut errors by 43% compared with the existing Scribe v2 Realtime. Both figures are based on evaluations across 102 languages.
The models have also been improved at distinguishing speakers in conversations involving multiple people. In recordings with three or more participants, Scribe v3 reduced speaker diarization errors by 17% compared with the previous model. Scribe v3 Live is the first to support real-time speaker diarization, and in ElevenLabs’ own evaluation it had 39% fewer speaker diarization errors than the comparison model.
The models have also improved at determining whether the other person has finished speaking in a real-time conversation. Whereas the previous approach determined whether an utterance had ended based on the duration of silence, the new model analyzes both the audio and the surrounding conversational context. It is designed to reduce instances of the AI interrupting when the other person pauses briefly.
The models also offer features tailored to the working environments of individual businesses. Scribe v3 allows users to register up to 1,000 brand names, abbreviations, technical terms and other items, and to specify formatting for dates, numbers and other details. The real-time model supports language changes during a conversation and also lets users pre-register frequently used technical terms.
ElevenLabs also shared examples of real-world business use. Dutch energy company Eneco tested Scribe v3 Live with a voice agent for customer service. As a result, the rate of correctly recognizing postal codes on the first try in customer calls increased by about 40%, while house-number recognition accuracy rose by about 15%.
Data security features for enterprise customers have also been strengthened. ElevenLabs provides independent data processing environments in the European Union (EU), India and Singapore. It also offers a “zero-retention mode” that permanently deletes audio and text data once a request has been processed. Businesses that need dedicated infrastructure can also deploy a private environment.
ElevenLabs plans to expand the supply of voice AI technology to South Korean businesses and developers. The company intends to help them use next-generation speech recognition models for tasks such as meeting transcription, customer service and voice agents.
[email protected] Jang Min-kwon Reporter