The reason talking artworks are mostly bad is latency. A person asks a sculpture a question, the sculpture thinks for two and a half seconds, and the social contract of conversation — already fragile when one party is a sculpture — collapses. Below about a second it starts to feel like being answered. Above it, it feels like waiting for a server.
ElevenLabs released two speech-to-speech models on 20 October 2026, and the headline figure is the one that matters for this: 0.57 seconds.
The two models
Eleven Realtime v1 Max is the larger one, aimed at more complex tasks such as customer support. It has a 0.57-second response time and scores 98.8% on Big Bench Audio. ElevenLabs’ claim is that “no model on the Artificial Analysis leaderboard is both faster and more accurate right now,” positioning it at the top-left of that leaderboard’s accuracy-against-latency plot.
Eleven Realtime v1 Mini is cheaper and aimed at natural conversation, and it has the feature we would actually reach for: it can tell when someone has finished talking or is just saying “mm-hmm.”
No latency or benchmark figures have been published for Mini.
Both are available through the ElevenAgents platform and through an OpenAI-compatible Realtime API. Pricing is per minute plus token costs, with Max dearer than Mini; no specific rates were given in the announcement.
Speech-to-speech is a different architecture, and it shows
The distinction that the announcement does not spell out is worth spelling out, because it is the reason the latency number is achievable at all.
The conventional voice pipeline is three models in a row: speech recognition turns your audio into text, a language model answers the text, and text-to-speech turns the answer into audio. Each stage waits for the one before it, each adds its own latency, and — the part that matters artistically — everything that was not a word gets thrown away at step one. Your hesitation, your rising tone, your sigh, the fact that you trailed off rather than stopped: all of it is discarded in the transcript, and then a synthesiser invents fresh prosody with no knowledge of yours.
Speech-to-speech skips the round trip through text. That is what buys the half second, and it is also what makes backchannel detection possible. “Mm-hmm” has no useful transcript. It is pure prosody — a signal that means keep going, I’m listening, carried entirely in pitch and duration. A pipeline that transcribes first cannot distinguish it from an interruption, which is why every voice assistant you have used stops talking when you agree with it.
A model that handles the audio directly can tell the difference. For an installation, that is the gap between a piece that converses and a piece that takes turns.
What to be sceptical about
The leaderboard claim is ElevenLabs’ own, made in a post on X, and has not been independently verified — including by us. Big Bench Audio at 98.8% is approaching the ceiling of that benchmark, which usually means the benchmark is close to exhausted rather than the problem close to solved.
The framing is also squarely commercial: customer support, agents, enterprise deployment. The announcement says nothing about languages, nothing about offline or on-premise operation, and gives no actual per-minute rate. For gallery or performance use, those three absences are the whole feasibility question — a piece that has to run for three months in a room with unreliable network, billed per minute of visitor attention, has a cost model nobody has published.
And the 0.57 seconds is a model response time, not a room-to-room time. Add microphone capture, network round trip, buffering and speaker output, and the number a visitor experiences will be meaningfully larger. It is still the best published starting point we have seen.
The one to watch is Mini
Max has the benchmark score; Mini has the behaviour. Knowing that a listener’s noise is encouragement rather than interruption is a small piece of social competence that no amount of accuracy substitutes for, and it is in the cheap model.
If you are building anything that talks to people in a room, that is the one to prototype with.
Sources
- Elevenlabs releases two new speech-to-speech models designed for real-time AI conversations — The Decoder
- Big Bench Audio release — Hugging Face Blog
- ElevenLabs pricing
- Artificial Analysis — Speech to Speech