ElevenLabs packages voice Agent: supports over 90 languages, compatible with OpenAI interface.

Beating AI News Flash: ElevenLabs has released its real-time speech model Eleven Realtime, comprising two versions: Mini and Max. Mini focuses on low-cost conversation, while Max targets customer service and complex tasks requiring continuous tool invocation. Both support over 90 languages and more than 17,000 voice styles.

In Artificial Analysis's 1,000 speech reasoning questions, Max achieved an accuracy rate of 98.8%, with an average latency of 0.57 seconds to start speaking. On the same benchmark, Qwen Audio 3.0 Realtime Plus had a slightly higher accuracy rate of 99.2%, but required an average of 1.54 seconds to start speaking.

Building a speech agent typically requires manually connecting speech recognition, large models, and speech synthesis, as well as tuning pauses, turn-taking, and interruptions. Eleven Realtime unifies and optimizes these components, eliminating the need for piecemeal assembly. Under the hood, it still chains multiple models together; when using Realtime mode, developers cannot select individual models within it.

The new models have begun rolling out through ElevenAgents and the Realtime API and are gradually being made available. The API is compatible with OpenAI's WebSocket protocol, allowing developers to reuse existing SDKs; projects using WebRTC will need to adjust their connection methods first.