Fish Audio Live 3 Answers in 186 Milliseconds, Then Spends the Call Deciding Whether to Speak

Fish Audio Live 3 is a full-duplex voice model that Fish Audio announced on 2026-11-10, and its entire pitch rests on one distinction most voice agents still get wrong: a pause in the middle of a phone number is not the same event as the pause at the end of a question. The model keeps listening while it speaks, can decide to respond in as little as 186 milliseconds according to Fish Audio, and was trained on recorded human conversation to learn when to acknowledge, when to wait, when to answer and when to stop — instead of being handed a table of silence thresholds.

That last clause is the interesting one, and it is also why this piece cannot give you a benchmark. Fish Audio has published no evaluation for Live 3, no per-byte price for it on its own rate card, and no model identifier or endpoint documentation naming it. Artificial Analysis, which scores duplex speech models on one harness, carries no Live 3 row as of 2026-11-11 — the board lists Fish Audio as a model family and scores none of its duplex work. The 186-millisecond number is vendor-reported and unaudited. What follows separates what Fish Audio has documented from what it has asserted, because one day after launch those are two different piles.

The pause that ruins a phone number

Fish Audio's own worked example is the best argument in the announcement. Read a number out in chunks — "415… 555… 0132" — and a conventional agent hears three separate end-of-turn events, or one long hesitation it mistakes for the caller thinking. Live 3 is documented to say "mm-hmm" or "okay, 415" between the chunks and then wait for the rest. The acknowledgement does two jobs at once: it confirms the digits landed, and it leaves the floor open so the caller can take a breath, check the next few digits and carry on.

The harder half of the same problem is the opposite error. If the agent is reading out opening hours and the caller says "yeah" or "mm-hmm" to show they are following, Live 3 is documented to keep going rather than stop. Then the caller says "Wait, actually, I just need Thursday," and it stops and makes room. Fish Audio's framing is that not every sound is an interruption, and the model is supposed to tell a backchannel from a takeover without a developer enumerating every sound a human might make.

The third case is corrections mid-utterance: "I'd like two pizzas… no, three. Three large. And… actually, make one of them a margherita." A pipeline that fires on the first plausible complete order places the wrong order twice. Live 3 is documented to wait through each revision. Support calls, bookings and orders are full of exactly this, and it is the reason Fish Audio's chief executive reached for a linguistic analogy rather than a latency number.

"The fix so far has been more rules and more thresholds," said Rissa Cao, Fish Audio's co-founder and CEO. "We think a model should learn this from real conversations, the same way language models learn grammar."

What Fish Audio has actually put on paper

Stripped of the marketing, the published surface of Live 3 is short enough to list in full. Everything below is from Fish Audio's announcement or its documentation; nothing here has been independently measured.

• Reaction time — Fish Audio states the model can decide to respond in as little as 186 milliseconds, in the context of short reactions such as "mm-hmm", "okay" and quick confirmations rather than full answers.

• Architecture — full duplex. It listens while it speaks, which is what makes a mid-number acknowledgement possible at all.

• Training signal — real human conversations, with turn-taking behaviour learned rather than configured. No dataset size, no hours, no evaluation set is disclosed.

• Languages — English only at launch, stated plainly in the announcement.

• Distribution — the Fish Agent Platform, Fish Audio's own API, and on-premises deployment for enterprise customers.

• Price — not published. Fish Audio's rate card lists drama-3, s2.1-pro, s2.1-pro-free, s2-pro and s1, plus three ASR models and voice-design-1. Live 3 does not appear on it.

The announcement itself is worth reading in full, because it is unusually concrete about behaviour and unusually quiet about numbers.

Three audio excerpts accompany it — a caller adding details with pauses, a caller cutting into a menu list, and a caller saying "mm-hmm" while the agent reads an order back — and each is the same behaviour claim from a different angle. None of them comes with a transcript-level score, and the announcement does not claim one.

What a Live 3 agent is made of

Live 3 is the speech layer of a larger product, and the rest of that product is where the integration work lives. Fish Audio runs a hosted Agent Platform, currently in public beta, which exposes a REST surface under /v1/agent/* plus client SDKs for JavaScript and React. Agents can be attached to a web widget, a phone number or a custom client, and sessions are authenticated with session tokens rather than a long-lived key in the browser. The platform handles the audio path; you bring the instructions, tools and workflows.

The part that matters for anyone building on top of it is the custom-LLM hook. Fish Audio lets an agent's reasoning layer point at an OpenAI-compatible endpoint, which means the model doing the thinking does not have to be the vendor's. That is a deliberate seam, and it is the seam where a routing layer belongs. OrcaRouter sits behind that endpoint with 216 models from the same key, at provider list price with zero markup and automatic failover when a provider errors — so the planner inside a Live 3 agent can be swapped, A/B-tested on real call transcripts and re-pointed without touching the speech layer or the phone number. A routing DSL decides which model gets which call, and model fusion lets a cheap model draft while a stronger one reviews the tool call before it executes. To be explicit about the boundary: OrcaRouter does not host Fish Audio Live 3. There is no third-party endpoint for it, and the speech layer here is the vendor's own.

The platform is where an agent gets its phone number, its tools and its session lifecycle.

What would settle it

Three numbers would turn this from a claim into a datapoint, and none of them exist yet.

• A per-byte or per-minute price for Live 3, so the cost of a ten-minute call can be compared with the cascade it replaces. Until that lands, any total-cost estimate for a Live 3 deployment is a guess with a decimal point.

• A latency distribution rather than a floor. "As little as 186 milliseconds" describes the fast tail of short reactions. Time-to-first-audio for a full answer is the number that decides whether a caller talks over the agent, and Fish Audio has not published it.

• An independent turn-taking score. Artificial Analysis runs Full-Duplex-Bench and a τ-Voice agentic test on duplex models, and both would be directly comparable to the published rows for the other duplex models in this market. A Live 3 row on either board would answer in one table what the announcement answers in prose.

In the meantime the honest read is this: the problem Fish Audio picked is real and under-served, the worked examples are the right ones, and the model is shipping to real users through a platform that is itself still in beta. If your agent's failure mode today is that callers get talked over, or that they have to repeat a digit sequence twice, Live 3 is aimed exactly at that — and the first thing to instrument after you deploy it is not latency but how often the caller has to start over.