Fish Audio Live 3: full-duplex voice model reacts in 186 ms

Fish Audio announced Live 3 on 10 November 2026, a voice model it describes as full duplex: it keeps listening while it speaks, instead of waiting for the caller to fall silent before it answers. According to Fish Audio, the model can decide to respond in as little as 186 milliseconds, a figure that applies to short reactions such as “mm-hmm”, “okay” or a quick confirmation rather than to a full answer. The company's own announcement page could not be read in full, so this article relies on the detailed account published by the routing service OrcaRouter and on the documentation changes that Fish Audio's team opened on GitHub.

The problem Live 3 tries to solve

Most voice agents still run on a chain of components: speech recognition, a language model, speech synthesis, and a set of silence thresholds that decide when the caller has finished. According to OrcaRouter's description of the announcement, that design confuses a pause in the middle of a phone number with the pause at the end of a question. Fish Audio's own example is a caller reading out “415… 555… 0132”: a conventional agent hears three separate end-of-turn events, whereas Live 3 is documented to say “mm-hmm” or “okay, 415” between the chunks and then wait for the rest.

The opposite error matters just as much. If the agent is reading out opening hours and the caller says “yeah” to show they are following, Live 3 is said to carry on; if the caller then says “Wait, actually, I just need Thursday”, it should stop and make room. A third case is self-correction, such as “two pizzas… no, three… actually, make one a margherita”, where an agent that acts on the first plausible order gets it wrong. Fish Audio presents these behaviours as learned from real human conversations rather than configured by hand.

The fix so far has been more rules and more thresholds. We think a model should learn this from real conversations, the same way language models learn grammar.

What is documented, and what is not

Here is what can be established at this stage, with no independent measurement behind any of it:

• Reaction time: as little as 186 ms, a figure supplied by Fish Audio, for short reactions.
• Language: English only at launch.
• Distribution: the Fish Agent Platform, Fish Audio's own API, and on-premises deployment for enterprise customers.
• Training: real human conversations; neither the size of the dataset nor the evaluation set is disclosed.
• Price: absent from the public rate card, according to OrcaRouter. A documentation pull request opened on GitHub on 10 November adds a Live 3 conversation row at $0.07 per minute, but it was still under review, and its author notes that the API fields (engine set to live3, and live3_prompt) were not yet in the OpenAPI schema and that the curl example had not been run.

There is no benchmark. According to OrcaRouter, Fish Audio has published no evaluation of Live 3, and Artificial Analysis, which scores duplex speech models, carried no Live 3 row as of 11 November. The 186 ms figure is therefore unaudited. It also describes the fast end of short reactions: the delay before the first sound of a full answer, which decides whether callers talk over the agent, has not been published. OrcaRouter itself sells model routing and presents its own product in the article, so its reading of the announcement should be weighed accordingly.

How an agent is built on it

Live 3 is the speech layer of the Fish Agent Platform, which is in public beta. It exposes REST endpoints under /v1/agent/ and SDKs for JavaScript and React; an agent can be attached to a web widget, a phone number or a custom client, with session tokens. According to the documentation pull request, Live 3 is switched on in the Builder or with voice.engine set to live3, situations are set through voice.live3_prompt, and users can switch back to the cascade mode. The same text compares the two modes on language support, turn-taking, delivery controls, transcripts and speech during warm transfers. The reasoning layer can point to an OpenAI-compatible endpoint, so the language model does not have to be Fish Audio's.

What it changes for a business

For a company that takes orders, bookings or support calls by phone in English, such as a restaurant chain, a clinic or a utility, the interest is practical: fewer callers cut off and fewer wrong orders. Before replacing an existing pipeline, it makes sense to wait for three things: a confirmed per-minute price, a latency distribution for full answers, and an independent turn-taking test. In the meantime, a pilot on a few hundred real calls, with interruptions and transcript errors measured and a way back to the old mode, keeps the risk low. Non-English markets will have to wait, since the model speaks English only.