Editorial illustration for Elevenlabs releases two speech-to-speech models for real-time AI conversations
Elevenlabs releases two speech-to-speech models for...
ElevenLabs put a number on its new speech models this week: 98.8 percent accuracy on the Big Bench Audio benchmark, with responses clocking in at 0.57 seconds. The company released two speech-to-speech models under the name Eleven Realtime, built specifically for AI systems that need to talk back and forth with people without the awkward pauses that give away a machine on the other end.
There are two versions. V1 Mini handles everyday conversation at a lower price point and is tuned to pick up on small verbal cues, like knowing the difference between someone wrapping up a sentence and someone just murmuring "mm-hmm" to show they're listening. V1 Max is aimed at heavier lifting, the kind of multi-step reasoning a customer support call might demand, where the model has to track context and respond accurately under pressure.
Both models run through ElevenLabs' ElevenAgents platform and plug into an OpenAI-compatible Realtime API, which should make them easier to drop into existing setups. Pricing works on a per-minute basis plus token costs, with Max carrying the higher price tag of the two. The company is making a specific claim about where this leaves the competition.
The smaller v1 Mini is built for natural conversations at lower cost and can tell when someone's done talking or just saying "mm-hmm." V1 Max targets more complex tasks like customer support.
Why this matters
For anyone building voice agents, the latency number is the headline here. A 0.57-second response time at 98.8 percent on Big Bench Audio puts Eleven Realtime Max in a different category than most speech-to-speech models we've tracked, where you're usually trading accuracy for speed or eating a full second of dead air before the model responds. That gap is what makes voice bots feel robotic, so closing it matters more than another few points on a benchmark leaderboard.
The Mini variant is the more interesting bet for founders watching margins: turn-taking detection that knows when someone's actually finished talking, at lower cost, is the difference between a usable customer-support agent and one that talks over callers. We'd want to see independent latency tests outside Elevenlabs' own Artificial Analysis comparison before taking the "fastest and most accurate" claim at face value. But if the numbers hold up under real traffic, this pushes real-time voice AI closer to something you'd actually want on a phone call, not just a demo.
Related Reading
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives Exa 3.0 Outperforms Rivals on 8 Benchmarks, Cuts Context Tokens by 95%
From Archives Barcelona Supercomputing Center Extends MareNostrum 5 for Spain-Based Startups
Further Reading
- Introducing Eleven Realtime: a simpler way to build natural voice agents - ElevenLabs
- Eleven Realtime | Speech to Speech Voice Agents - ElevenLabs
- ElevenLabs stellt Eleven Realtime vor: Speech-to-Speech-Modelle mit unter einer Sekunde Latenz - The Decoder
- Eleven Realtime v1 Mini vs Max: 570ms, 98.8% Big Bench Audio - ExplainX
