ElevenLabs Unveils Packaged Voice Agent Supporting Over 90 Languages, Compatible with OpenAI API

Dongcha Beating AI Flash News: ElevenLabs has launched Eleven Realtime, a real-time voice model available in two versions: Mini and Max. Mini is optimized for low-cost conversational scenarios, while Max is tailored for customer service and complex tasks requiring continuous tool integration. Both models support over 90 languages and more than 17,000 voice profiles. In Artificial Analysis’s 1,000-voice inference benchmark, Max achieved a 98.8% accuracy rate, with an average response time of 0.57 seconds. On the same leaderboard, Qwen Audio 3.0 Realtime Plus posted a slightly higher accuracy of 99.2%, but its average time to first response was 1.54 seconds. Building a voice agent usually requires developers to separately connect speech recognition, large language models, and text-to-speech systems, plus debug pauses, turn-taking, and interruption handling. Eleven Realtime unifies and optimizes these components, removing the need for piecemeal integration. While the underlying architecture still relies on multiple connected models, developers cannot select individual models when using Realtime mode. The new model is now being rolled out via ElevenAgents and the Realtime API, with access gradually expanding. The API is compatible with OpenAI’s WebSocket protocol, enabling developers to reuse existing SDKs; projects using WebRTC will need to adjust their connection methods first.