ElevenLabs launched Eleven Realtime, a new class of Speech-to-Speech models for voice agents. It offers two models: v1 Mini for low-cost natural conversations and v1 Max for complex tasks. Both support 90+ languages and 17,000+ voices from the ElevenLabs library. The models are available in ElevenAgents and via the Realtime API. The Realtime API is compatible with OpenAI’s Realtime API. Eleven Realtime v1 Max scored 98.8% on Big Bench Audio for speech reasoning. v1 Mini scored 75.9% on the same benchmark. v1 Max starts speaking in 0.57 seconds on average. v1 Mini starts speaking in 0.62 seconds on average. v1 Mini scored 95.8% on Full-Duplex-Bench for conversational dynamics. v1 Max scored 94.1% on Full-Duplex-Bench. v1 Max achieved a 998 score in the Speech Agent Arena. v1 Mini achieved a 988 score in the Speech Agent Arena. v1 Max completed 87.1% of tool-use tasks. v1 Mini completed 60.3% of tool-use tasks.
What changed
Building a voice agent required assembling and tuning separate components like speech recognition, voice activity detection, turn-taking, LLM, and text-to-speech.
What it unlocks
Deploying a complete, natural-sounding voice agent without assembling separate components.
- 98.8% correct on Big Bench Audio (v1 Max)
- 75.9% correct on Big Bench Audio (v1 Mini)
- 0.57s average time to first audio (v1 Max)
- 0.62s average time to first audio (v1 Mini)
What you need to act on it
- ElevenAgents or Realtime API access
- WebSocket integration for API use
Sources