Voice and NPCs
Fish Audio Live 3: a full-duplex voice model with 186 ms replies
Fish Audio has released Live 3, a voice model that keeps listening while it speaks and can start a reply in as little as 186 milliseconds. It is available by API at $0.035 a minute, which makes natural turn-taking cheap enough to try in a talking character.
On 10 November, Fish Audio introduced Live 3, a full-duplex voice model. It keeps listening while it speaks, so a pause, a quick “mm-hmm” or a change of mind can shape what it does next.
Most voice agents chain separate parts: speech recognition, a language model, speech synthesis and a rule-based turn-taking layer. Fish Audio says that stack is where conversations go wrong. An agent can answer quickly and still speak at the wrong moment. Live 3 moves conversational timing into the speech model itself.
What it does
The company trained Live 3 on real human conversations, so it learns when to acknowledge, wait, respond or stop. The post says these behaviours were learned from people, not written as a list of rules. It describes three situations:
- Acknowledging without taking over. When a caller reads out a phone number in chunks, Live 3 can say “mm-hmm” or “okay, 415” between them and wait for the rest.
- Telling a “yeah” from an interruption. If you say “yeah” while it explains opening hours, it carries on. If you say “Wait, actually, I just need Thursday”, it stops.
- Waiting through corrections. In an order such as “two pizzas… no, three… actually, make one a margherita”, it holds back instead of confirming the wrong item.
Fish Audio publishes three audio excerpts from real test calls, with the original pauses and overlaps left in.
The numbers and access
- Latency: Live 3 can decide to respond in as little as 186 milliseconds. Short reactions fill the gap while the full answer is prepared.
- Language: English only for now.
- Access: the Fish Agent Platform and the API. Enterprises can deploy it on-premises.
- Price: $0.035 a minute, which the product page lists as a limited-time price against a standard $0.07. Phone calls add $0.01 a minute. The product page lists a $5 starter credit.
- Voices: the platform offers more than two million voices in over 30 languages, including your own clone, though the post does not say which of them Live 3 uses.
Limits
The announcement gives no parameter count, no open weights and no published benchmark. The 186 ms figure is a best case, not a median, and the post does not say how it was measured. The product page is aimed at customer-service agents (bookings, refunds, CRM hand-offs), so anything for games is your own adaptation. Treat the demos as the evidence and test it with your own scripts.
Our take
For people building interactive worlds, this is the third full-duplex voice release in a few days, after Higgs Realtime Duplex and Shizuku-Duplex. The common idea is that turn-taking is learned, not scripted. That matters for a character who must not talk over the player.
What stands out here is the price and the packaging. An API at a few cents a minute lets a small team prototype a talking shopkeeper or guide this week. The gap is character: Live 3 is built around instructions, tools and workflows, so the personality, memory and world knowledge still have to come from your own design. We look at that side in intelligent NPCs.