OPEN LINK ↗
// 1h agoMODEL RELEASE
Eleven Realtime Launches 570ms Voice Agents
Eleven Realtime combines speech recognition, turn-taking, an LLM, and voice into end-to-end speech-to-speech models: v1 Mini for fast conversations and v1 Max for complex tasks. ElevenLabs reports 620ms and 570ms to first audio, 90+ language support, and an OpenAI-compatible WebSocket API. [Announcement](https://elevenlabs.io/blog/realtime) [Docs](https://elevenlabs.io/docs/overview/models/)
// ANALYSIS
ElevenLabs is targeting voice agents’ biggest weakness—latency and awkward handoffs—with a managed, tuned stack rather than another standalone speech model. The deeper move is making the entire voice pipeline a model-selection decision while preserving a familiar API path.
- –End-to-end tuning could reduce brittle orchestration across ASR, turn-taking, LLMs, and TTS.
- –v1 Max’s reported 98.8% Big Bench Audio score and 87.1% tool-use completion position it for support, scheduling, and workflow agents.
- –OpenAI compatibility lowers migration friction, though WebRTC integrations must move to WebSocket.
- –The tradeoff is control: teams needing specific LLM, ASR, or TTS components may prefer ElevenLabs’ custom mode.
- –Per-minute billing plus token costs means developers still need to benchmark quality against total call economics.
// TAGS
eleven-realtimevoice-agentspeechttssttapistreaming
DISCOVERED
1h ago
2026-10-23
PUBLISHED
7h ago
2026-10-23
RELEVANCE
9/ 10