ElevenLabs Launches Scribe v3 Live to Stop Voice Agents Cutting Off Callers

At a glance

ElevenLabs released two speech recognition models, Scribe v3 and the real-time Scribe v3 Live. They are built to understand speech well even when people talk over each other or there is background noise.

Key points

1. 1

   The company says Scribe v3 cuts word errors by about 25% compared with its predecessor across 102 languages.

2. 2

   Independent benchmarks rank its accuracy near the top, but it is slower and more expensive than models from Microsoft and xAI.

3. 3

   Scribe v3 Live checks conversation context during brief pauses to judge whether a speaker has really finished.

4. 4

   Users can preload specialized terms like brand names and medications, and enterprise customers get EU-only servers and a mode that deletes data right after each request.

Why it matters

It addresses voice agents cutting callers off, which matters to companies and developers automating phone support.