Fish Audio Releases Live 3, a Full-Duplex Voice Model That Reacts in 186 ms
Nov 10, 2026, 8:43 PM · MarkTechPost
Fish Audio is betting voice agents fail on timing, not speed, and Live 3 tries to learn when to talk from real conversations. The pitch is right; the proof isn't published yet.
Why it matters
Anyone who has used a phone bot knows the failure: it cuts you off mid-number or keeps talking after you try to interrupt. Fish Audio's Live 3 is a full-duplex model that listens while it speaks and moves turn-taking from hand-tuned rules into the speech model itself, trained on real human conversations.
It's cheap, too: $0.035 a minute until December 31, then a $0.07 list rate that covers speech recognition, speech synthesis, basic language models and tool calls. Pricing like that pulls voice agents into more call centers and small businesses.
From the desk
We think Fish Audio has the problem framed correctly. Replying fast and replying at the right moment are different skills. Their examples make the case: briefly acknowledging chunks of a phone number and then waiting, telling a casual yes apart from a real interruption, and holding off while a caller corrects a pizza order instead of logging the wrong one. Silence thresholds can't tell those apart. A model trained on how people actually talk might.
But there are no benchmark scores. Fish Audio shows demos and a 186-millisecond figure, and that number is a floor for quick acknowledgments, not the time to a full answer. There's no latency distribution, and it's English only. Rivals in the comparison either publish turn-taking scores or cover dozens of languages. Until independent numbers arrive, treat Live 3 as a promising design, not a proven leader.
The downside to name is the one that comes with every better voice agent. The platform advertises more than two million library voices, including cloned ones. More natural timing makes bots harder to tell from people on the phone, which helps customer service and also helps scammers. As voice agents get cheaper and smoother, disclosure that you're talking to a machine, and consent around cloned voices, matter more, not less.
Our read: useful step, honest problem statement, missing receipts. Developers should test it on their own call flows before believing any demo.
Context
Live 3 runs on Fish Audio's Agent Platform, which is in public beta, through its API, and on-premises for enterprises. Agents can answer phone numbers, connect over WebRTC or sit in a website widget; phone calls add $0.01 per minute.
Who feels it
- Developers
- Turn-taking logic moves into the model, which could cut pipeline tuning, but there are no published scores to plan around.
- Contact centers
- Low per-minute pricing makes pilots cheap; test interruption and correction handling on real calls first.
- Consumers
- More human-sounding bots raise the stakes for clear AI disclosure and limits on voice cloning.
What to watch
- Independent turn-taking and latency results, such as an Artificial Analysis score
- Support for languages beyond English
- Pricing and adoption after the promotional rate ends December 31