ElevenLabs Launches Eleven Realtime, a Voice Model That Responds in 0.57 Seconds

At a glance

ElevenLabs has released Eleven Realtime, a speech-to-speech (STS, a method that turns voice directly into voice) model that combines speech recognition, conversation flow control, and voice synthesis in one model. It supports more than 90 languages, including Korean. It comes in two versions: Mini for fast conversation and Max for complex tasks.

Key points

1. 1

   Developers no longer need to pick and connect separate technologies to build voice agents. They can choose one model instead.

2. 2

   Max starts responding 0.57 seconds after a question ends and scored 98.8% on a voice-based reasoning test (the ability to listen to speech and make judgments).

3. 3

   Mini scored 95.8% on a test measuring interruptions and turn-taking, higher than Max's 94.1%.

4. 4

   It works with existing voice agent platforms and is compatible with the OpenAI Realtime API, making it easy to switch.

Why it matters

This could change how companies and developers build voice customer service bots, including their production methods and costs. However, pricing has not been disclosed.