Live 3

Live 3 is the full-duplex voice model Fish Audio released on 10 November 2026 for real-time voice agents. It keeps listening while it speaks, so a pause, a short acknowledgment, an interruption or a mid-sentence change of mind can shape the next turn. The official write-up is the blog post Introducing Live 3: Built for the rhythm of conversation. The product page is Live 3, and a try-it entry point is fish.audio/agents. The official account @FishAudio posted at 2026-11-10T18:02:05Z; co-founder and CEO Rissa Cao’s launch post is 2120214987633467626.

The official blog separates being fast from having a sense of timing: an agent can start a reply quickly and still speak at the wrong moment. A pause after “What time do you open?” does not mean the same thing as a pause in the middle of a phone number. Most voice agents split speech recognition, reasoning, speech generation and turn-taking across a pipeline, then stitch those parts together with rules. Live 3, Fish Audio says, learns conversational timing inside the speech model itself, from real human conversations, so it can learn when to acknowledge, wait, respond or stop. “The fix so far has been more rules and more thresholds,” Cao said in the post. “We think a model should learn this from real conversations, the same way language models learn grammar.”

The blog gives concrete behaviors. If a caller reads a number in chunks — “415… 555… 0132” — Live 3 might say “mm-hmm” or “okay, 415” and then leave the floor open so the person can breathe and continue. Officially, it can decide to respond in as little as 186 milliseconds; short reactions such as “mm-hmm,” “okay,” and quick confirmations keep the exchange moving while a full answer is prepared. If a user says “yeah” or “mm-hmm” while the agent is explaining hours or services, it keeps going. If the user says “Wait, actually, I just need Thursday,” it stops and takes the new request. When someone revises an order — “two pizzas… no, three. Three large. And… actually, make one of them a margherita” — the company says Live 3 waits through each correction instead of jumping in with the wrong item. The post includes three excerpts from test calls: finishing an order, cutting in with a request, and acknowledging while an order is read back.

On availability, the official blog says Live 3 speaks English today, is available on the Fish Agent Platform and through the API, and can be deployed on-premises for enterprises. The product page places it in a platform for building, deploying and monitoring voice agents that sound human and get work done, and lists live monitoring, turn-by-turn transcripts, latency breakdowns, automatic summaries and success criteria. The page shows promotional pricing that includes Live 3 at $0.035 per minute, 50% off a $0.07 list price; whether that promo is still on, and how full-duplex versus cascaded usage is billed, should be checked on the live site. Cao’s post also offered a demo for enterprises and a free custom agent for the first 1,000 people who replied with a use case — a launch offer, not necessarily a standing term.

Key links:

• Blog: https://fish.audio/blog/introducing-live-3/ ; product: https://fish.audio/live3/ ; try it: https://fish.audio/agents ;
• Official post: https://x.com/FishAudio/status/2120215430807904395 ; CEO post: https://x.com/rissa_cao/status/2120214987633467626 ;
• This is a specific full-duplex voice-model name from Fish Audio. Do not merge it with the company’s text-to-speech models (for example Drama 3 or S2.1 Pro), or with the generic phrase “live agent.”

Why it matters

Voice agents often fail at turn-taking: they interrupt, they treat a backchannel as a barge-in, or they lock in the wrong order when the caller changes their mind. Live 3 turns that problem into a named, callable model, with an official blog, a product page, API and platform access, and an on-prem option — not only a demo video. For teams building phone support, booking or ordering agents, that is a new name you can try now. The official 186-millisecond figure and the promotional per-minute price are company numbers and should be cited as such.

The launch post went out at 2026-11-10T18:02Z (02:02 on 11 November, Asia/Shanghai). Earlier slots at 04:32, 08:32 and 12:32 in this timezone missed it; this slot records it. English is the only language officially listed today.

Limits

The call excerpts in the blog are official tests, not a third-party evaluation. Pricing, quotas and the free custom-agent offer should be read from the live site and official posts. Do not use Live 3 as a catch-all name for every Fish Audio speech product.

1. Read the official blog https://fish.audio/blog/introducing-live-3/ and the product page https://fish.audio/live3/ , and cross-check the announcement at https://x.com/FishAudio/status/2120215430807904395 ;
2. Try it at https://fish.audio/agents , or integrate through the Fish Agent Platform or API as the product page describes; for on-premises use, follow the official documentation or book a demo ;
3. When citing full duplex, 186 milliseconds, or $0.035 per minute, say the source is the 10 November 2026 official blog or product page, and re-check the live price and terms ;
4. Do not treat Live 3 as Fish Audio’s TTS models (such as Drama 3 or S2.1 Pro) or as a generic “live agent”; this page only restates the official blog, product page and announcement — re-check the upstream source before citing it.