ElevenLabs on November 3 unveiled "Eleven Realtime," a speech-to-speech model that integrates speech recognition, conversational flow control, and speech synthesis into a single model selection, offered in two versions: "Eleven Realtime v1 Mini," focused on fast and natural conversation, and "Eleven Realtime v1 Max," specialized for complex tasks, available through the "Eleven Agents" platform and a Realtime API. Both models support more than 90 languages, including Korean and Japanese, and let users select a voice from a library of more than 17,000 voices, with features including turn-taking, responsiveness to user interruptions, and the ability to call external tools. The Mini model is suited to AI assistants, educational tutors, and game and entertainment characters, while the Max model focuses on customer support, lead qualification, and multi-step tool-calling tasks. In an evaluation conducted last October by third-party evaluator Artificial Analysis, the Max model recorded 98.8% accuracy on "Big Bench Audio," with 0.57 seconds from when a question ended to when the answer began; on the "Full Duplex Bench," the Mini model scored 95.8% and the Max model 94.1%.