ElevenLabs says Scribe v3 cuts word errors by about 25 percent
Scribe v3 is behind only 1 preview model in word errors, but 2 rivals are many times as fast and cost half.
ElevenLabs released 2 models, Scribe v3 and its Live version, that change voice to text. They are for audio with noise and for persons who speak at the same time. Scribe v3 is for recorded files. The Live version works in real time. It uses the context of the call to know when a caller is done. An agent then does not cut off a caller who stops for a short time.
How it works
Many voice agents use pause timers of a set length. They cut off callers who stop to think or who read numbers slowly. The Live model reads the words of the call and the audio to know when a speaker is done. Developers can give it thousands of special terms, such as names of products or medicines.
The numbers
ElevenLabs says word errors fall about 25 percent against the previous model in 102 languages. Artificial Analysis, a test group that is not connected to ElevenLabs, says only a preview model from Alibaba is more accurate. Models from Microsoft and xAI are many times as fast and cost half.
What is not known
The test measures only word errors on recorded files. It does not test speaker labels, rare languages or the work of the Live model in real time. ElevenLabs gives no number for how frequently agents cut off callers.
Insight: ElevenLabs sells the full voice stack and each part
In October 2026, ElevenLabs launched Eleven Realtime, one model that replaces the different parts for voice recognition, language and voice output. At this time it also improves the recognition part, and adds partners and a marketplace. This seems to be a plan to get a buyer at each layer, as a full stack or in parts. Test if you can change one part without new work.
Sources
Posted
