ElevenLabs launches two voice agent models, the top one scores 98.8%
In a test by Artificial Analysis, Eleven Realtime v1 Max gave 98.8% correct answers to 1,000 spoken reasoning questions.
ElevenLabs launched Eleven Realtime, a new class of speech to speech models. Each model is a complete voice agent. A team can use one model. It does not assemble different parts for speech recognition, the language model and voice output. The two models are v1 Mini for low cost and v1 Max for complex tasks. Each model works in more than 90 languages.
How it works
Mini is for assistants, teachers and interactive characters. Max is for customer support, lead qualification and workflows with many steps. The API is compatible with the Realtime API of OpenAI. A team changes the endpoint and the model name in its code.
The numbers
Artificial Analysis tested the two models. On Big Bench Audio, Max gave 98.8% correct answers and Mini gave 75.9%. Max starts to speak in 0.57 seconds. On tool use, Max completed 87.1% of tasks and Mini completed 60.3%.
What is not known
The Realtime API uses WebSocket only. A WebRTC integration must change to WebSocket first. The API runs apart from ElevenAgents, thus channels and a knowledge base are in ElevenAgents only. The source gives no price.
Insight: ElevenLabs makes it easy to change in and hard to stop
ElevenLabs made its API compatible with the API of OpenAI. A team changes only the endpoint and the model name. Phone, WhatsApp, guardrails and tests stay in ElevenAgents. In a test by Artificial Analysis, Fish Audio had a higher pronunciation score than Eleven v4. This reads as a plan to win teams with an easy change, then to keep them with the platform. Check what you give up if you stop.
Sources
Posted
