Voice AI has a timing problem. Most voice agents can reply fast — but they speak at the wrong moment. A pause after “What time do you open?” usually ends a turn. A pause in the middle of a phone number does not. Traditional voice agents handle this with hand-tuned rules and silence thresholds, which break the moment a caller speaks naturally.

Fish Audio’s Live 3 takes a different approach: it moves turn-taking into the speech model itself. The model listens and speaks simultaneously (full-duplex), learning when to acknowledge, wait, respond, or stop from real human conversations. Short reactions like “mm-hmm” can start in as little as 186 ms.

This article covers what Live 3 is, how it handles turn-taking, deployment options, pricing, and how it compares to OpenAI GPT-Live-1, Eleven Realtime v1 Max, and Gemini 3.8 Live.

Disclosure: This article contains no affiliate links. All data is sourced from Fish Audio’s official documentation, launch post, and third-party trackers.

What is Fish Audio Live 3?

Live 3 is a full-duplex voice model from Fish Audio, built for real-time voice agents. “Full-duplex” means it listens and speaks at the same time — like a human conversation. Traditional voice agents are “half-duplex”: they listen, then speak, then listen again, using silence detection to decide when a turn ends.

The core argument from Fish Audio: speed and timing are different problems. An agent can reply fast and still speak at the wrong moment. Live 3 solves timing by learning it from data rather than programming it with rules.

Live 3 runs on:

• Fish Agent Platform (public beta) — REST API under /v1/agent/*
• Fish Audio API — direct API access
• On-premises — for enterprises that need full control

Browser sessions run through @fishaudio/agent-client or @fishaudio/agent-react. Agents can answer phone numbers, connect over WebRTC, or sit in a website widget. The platform includes 2,000,000+ library voices, including cloned voices.

How does Live 3 handle turn-taking?

Most voice agents split recognition, reasoning, speech generation, and turn-taking across separate pipeline stages. Fish Audio states that Live 3 moves conversational timing into the speech model itself. It was trained on real human conversations to learn when to acknowledge, wait, respond, or stop.

Fish Audio’s team walks through three behaviors:

1. Acknowledging without taking over

A caller reads a phone number: “415… 555… 0132.” Live 3 may say “okay, 415” between chunks, then wait for the rest. It acknowledges without seizing the turn.

2. Telling a backchannel from an interruption

A caller saying “yeah” lets the agent continue speaking. “Wait, actually, I just need Thursday” makes it stop. The model distinguishes between supportive backchannels and genuine interruptions.

3. Waiting through corrections

For “two pizzas… no, three,” Live 3 waits instead of logging the wrong order. It doesn’t jump on the first thing it hears.

Fish Audio frames the old approach as piling on “more rules and more thresholds.” Live 3 instead learns timing from data. A separate reasoning layer still holds the agent’s instructions and tools.

How fast is Live 3?

Fish Audio states Live 3 can decide to respond in as little as 186 ms. This figure covers short reactions such as “mm-hmm,” “okay,” and quick confirmations. The full answer is prepared in parallel.

Important distinction: 186 ms is a floor for small reactions, not time-to-first-audio for a complete reply. Fish Audio has not published a latency distribution or average response time. Third-party tracker OrcaRouter also reports no Artificial Analysis score for Live 3 as of November 11.

What we know:

• Minimum decision time: 186 ms (short reactions)
• Full answer preparation: runs in parallel with the decision
• Latency distribution: not published
• Artificial Analysis score: not available yet

Deployment options

Live 3 is available in three deployment modes:

• Fish Agent Platform (public beta) — hosted, REST API under /v1/agent/*. Browser sessions via @fishaudio/agent-client or @fishaudio/agent-react.
• Fish Audio API — direct API access for custom integrations.
• On-premises — for enterprises that need full control over their data and infrastructure.

Agents can answer phone numbers, connect over WebRTC, or sit in a website widget. Fish Audio’s current docs note that with Live 3 enabled (voice.engine: "live3"), speech recognition is fixed to fish_audio:transcribe-stream-1, which also classifies overlaps as interruptions or acknowledgements.

Pricing

Pricing note: The $0.035/min promo rate is available until December 31, 2026. After that, the list rate of $0.07/min applies. The rate covers speech-to-text, text-to-speech, basic LLM inference, knowledge retrieval, and tool calls — all bundled.

How does Live 3 compare to other full-duplex voice models?

ElevenLabs scores come from Artificial Analysis, as reported by ElevenLabs. OpenAI and Google figures are self-reported. Fish Audio has not published benchmark scores.

Key takeaways

• Live 3 is Fish Audio’s full-duplex voice model for phone and web agents.
• It learns turn-taking from real conversations instead of silence thresholds and hand-tuned rules.
• Short reactions can start in as little as 186 ms — a floor for small reactions, not full-response time.
• Agents cost $0.035 per minute until December 31, 2026, then $0.07 (covers STT, TTS, basic LLM, knowledge retrieval, tool calls).
• English only — no multilingual support yet.
• No published benchmark scores — no Artificial Analysis score, no latency distribution. This is the biggest gap vs. competitors.
• Full-duplex design means the model listens and speaks simultaneously, handling backchannels, interruptions, and corrections naturally.

Who should use Fish Audio Live 3?

Good fit:

• Developers building phone or web voice agents who need natural turn-taking
• Teams that want an all-in-one voice stack (STT + LLM + TTS) at a predictable per-minute price
• English-language use cases (customer support, sales, scheduling, FAQ)
• Enterprises needing on-premises deployment

Not a good fit (yet):

• Multilingual applications (English only)
• Teams that need published benchmark scores for evaluation
• Use cases requiring sub-100ms latency guarantees (no latency distribution published)
• Teams already invested in OpenAI or ElevenLabs ecosystems

Frequently asked questions

What is Fish Audio Live 3?
A full-duplex voice model from Fish Audio that listens and speaks simultaneously. It learns turn-taking from real human conversations instead of silence thresholds. Short reactions can start in as little as 186 ms.

How much does Fish Audio Live 3 cost?
$0.035 per minute (promotional rate until December 31, 2026, down 50% from $0.07). The rate covers speech-to-text, text-to-speech, basic LLM inference, knowledge retrieval, and tool calls. Phone calls add $0.01/min. Phone numbers cost $1.20/month. New accounts get a $5 trial balance.

Is Fish Audio Live 3 available for self-hosting?
Yes. Live 3 is available on the Fish Agent Platform (public beta), through the Fish Audio API, and on-premises for enterprises.

How does Fish Audio Live 3 compare to OpenAI GPT-Live-1?
Live 3 learns timing in the speech model and is cheaper ($0.035 vs $0.05/min). GPT-Live-1 delegates reasoning to a backend model, supports telephony natively, and self-reports +30 pp on Full Duplex Bench. Live 3 is English-only; GPT-Live-1’s language support is undisclosed.

How does Fish Audio Live 3 compare to Eleven Realtime v1 Max?
Eleven Realtime v1 Max is cascaded (not full-duplex in the same way) but supports 90+ languages and has published benchmark scores (94.1% Full-Duplex-Bench). Live 3 is full-duplex with learned timing but English-only with no published scores. ElevenLabs is more mature; Fish Audio is cheaper and more focused on timing.

Does Fish Audio Live 3 support multiple languages?
No. Live 3 is English-only as of November 2026. Multilingual support has not been announced.

Has Fish Audio Live 3 been benchmarked?
No. Fish Audio has not published benchmark scores (no Artificial Analysis score, no latency distribution). The only published metric is the 186 ms minimum decision time for short reactions. Third-party tracker OrcaRouter also reports no Artificial Analysis score.

Conclusion

Fish Audio Live 3 is an interesting entry in the full-duplex voice model space. Its core innovation — learning turn-taking from real conversations instead of silence thresholds — addresses a real problem that plagues most voice agents: speaking at the wrong moment. The 186 ms minimum decision time for short reactions is impressive, and the $0.035/min pricing (until Dec 31, 2026) is competitive.

But the lack of published benchmark scores, latency distributions, and multilingual support are significant gaps. OpenAI, ElevenLabs, and Google all have more mature offerings with published metrics. Fish Audio’s bet is that learned timing beats programmed timing — and that bet is worth watching, but unproven at scale.

If you’re building English-language voice agents and want to try a full-duplex model with natural turn-taking, Live 3 is worth testing. The $5 trial balance makes it low-risk. But if you need multilingual support, published benchmarks, or enterprise-grade reliability, the established providers (OpenAI, ElevenLabs, Google) are safer choices today.

Try it at fish.audio/live3. Check the Agents docs for integration details.