Introducing Live 3
- Source
- fish.audio
- Author
- Fish Audio editorial sitemap
- Date

Context
Fish Audio's Live 3 is a voice meant to fix when a phone talks rather than how fast it answers, by listening the whole time it is speaking (Fish Audio). Fish Audio calls this full-duplex, meaning sound can travel both ways at once, as in a normal phone call. The company's point is that a quick reply can still come at the wrong moment. A pause after "What time do you open?" usually ends a question, while a pause halfway through reading out a phone number means more is coming. Most voice agents, the post says, coordinate speech recognition, reasoning, speech generation and turn-taking as separate parts of a pipeline, and teams tune that timing with added rules and thresholds. Fish Audio instead trained Live 3 on real human conversations, so the model itself learns when to acknowledge, wait, respond or stop.
The examples in the post show three behaviors. When a caller reads a number in chunks, Live 3 can answer with "mm-hmm" or repeat the digits and then wait for the rest. When a caller says "yeah" while the agent is explaining, it keeps talking, but when the caller says "Wait, actually, I just need Thursday," it stops and takes the new request. When someone revises a pizza order mid-sentence, it waits through each correction instead of confirming the wrong item. Fish Audio says the model can decide to respond in as little as 186 milliseconds, and that short acknowledgments fill the gap while the full answer is prepared. The post gives no test conditions for that figure and no comparison with other systems, and its evidence for the timing behavior is three recorded excerpts from test calls.
Live 3 handles English only for now. It is available on Fish Audio's agent platform and through its , and enterprises can run it on their own servers. The developer still supplies the instructions, tools and workflows; Live 3 supplies the conversational timing. It joins other full-duplex voice models such as OpenAI's GPT-Live-1, so the claim to test is not that Live 3 can listen while speaking but whether its learned timing handles callers' pauses and corrections better.
- Fish Audio argues response speed and conversational timing are separate problems: an agent can reply fast and still speak at the wrong moment, because a pause after a question and a pause mid phone number mean different things.
- When a caller reads out a number in chunks ("415… 555… 0132"), Live 3 can interject "mm-hmm" or "okay, 415" between them, confirming the digits were heard while leaving the floor open for the rest.
- On self-corrected orders ("two pizzas… no, three… make one a margherita"), Live 3 waits through each revision instead of jumping in with the wrong item; Fish Audio says support calls, bookings and orders are full of these moments.
- Live 3 is English-only for now and is available on the Fish Agent Platform, through the API, and for on-premises enterprise deployment.
- Model — A computational system that uses learned patterns to produce an output from information supplied to it.
- Agent — A system in which a model repeatedly chooses operations, a harness carries them out, and the results return to the model to inform its next choice.
- API — Application programming interface: the agreed set of operations, inputs and outputs through which one piece of software uses another.
Checking sign-in…
Loading comments…






