ElevenLabs voice model responds in 0.57 seconds

The time it takes a voice AI to hear a question and answer it (response latency) has dropped to 0.57 seconds. ElevenLabs unveiled 'Eleven Realtime,' a model that bundles speech recognition, conversation flow control, and speech synthesis into one, on November 3 (local time), revealing the figure (measured on the Max version) recorded in October by AI evaluation firm Artificial Analysis.

ElevenLabs released its voice-based conversational agent model 'Eleven Realtime' in two versions: 'v1 Mini,' tailored for AI assistants, educational tutors, and game characters where fast conversation matters, and 'v1 Max,' specialized for complex tasks such as customer support and lead qualification. Both models support more than 90 languages, including Korean and Japanese, and come with barge-in responsiveness, external tool calling, and natural turn-taking in conversation. Voices can be selected from the company's library of more than 17,000 voices.

Benchmark scores vary by model. On 'Big Bench Audio,' which measures voice-based reasoning ability, Max recorded 98.8% accuracy. On 'Full Duplex Bench,' which measures silence handling, turn-taking, and barge-in responsiveness, Mini scored 95.8% versus Max's 94.1%, putting Mini 1.7 percentage points ahead. However, only the Max figures were disclosed for Big Bench Audio accuracy and time to first response.

The company also lowered barriers for developers. The realtime API is compatible with OpenAI's Realtime API and connects via WebSocket. Developers already using the OpenAI API can switch by simply changing the endpoint and model name. WebRTC is not supported at launch, however. The model is also available through the voice conversational agent platform 'Eleven Agents,' where users can select it directly when creating a new agent or from the model selection menu of an existing one. Deployment across phone, website, and contact center channels allows existing workflows, knowledge bases, and safety guardrails to be used as-is.

The company also signaled its intent to target the Korean market. A company official said, "With Eleven Realtime supporting more than 90 languages including Korean, we aim to help domestic companies and developers build voice agents that converse naturally with customers." However, Eleven Realtime's pricing structure was not disclosed in this announcement.