Fish Audio has released Live 3, a full-duplex voice model capable of responding in as little as 186 ms. The system allows callers to pause, offer backchannel noise like “mm-hmm,” or correct themselves without breaking the flow of a call. Fish Audio states that this capability moves turn-taking logic from rigid pipeline rules into the speech model itself.
How does Live 3 handle turn-taking?
Traditional voice agents typically split recognition, reasoning, and generation into separate stages with hard-coded thresholds. The Fish Audio team argues that speed and timing are distinct problems. An agent can reply quickly yet speak at the wrong moment. A pause after a question like “What time do you open?” usually ends a turn, whereas a pause in the middle of a phone number does not.
Live 3 was trained on real human conversations to learn when to acknowledge, wait, respond, or stop. The developers describe older approaches as piling on “more rules and more thresholds.” Instead, the model learns timing from data, while a separate reasoning layer retains the agent’s instructions and tools.
The team demonstrates three specific behaviors:
- Acknowledging without taking over. If a caller reads “415… 555… 0132,” Live 3 may say “okay, 415” between chunks and then wait.
- Telling a backchannel from an interruption. A caller saying “yeah” lets the agent continue. “Wait, actually, I just need Thursday” makes it stop.
- Waiting through corrections. For “two pizzas… no, three,” Live 3 waits instead of logging the wrong order.
How fast is Live 3?
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.
This is a floor for small reactions, not time-to-first-audio for a complete reply. Fish Audio has not published a latency distribution. Third-party tracker OrcaRouter also reports no Artificial Analysis score for Live 3 as of November 11.
How do developers deploy Live 3?
Live 3 is available on the Fish Agent Platform, through the Fish Audio API, and on-premises for enterprises. The Agents platform is currently in public beta.
Agents expose a REST surface under /v1/agent/*. 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 Live 3 page lists 2,000,000+ library voices, including cloned voices.
A pending docs pull request shows Live 3 enabled with voice.engine: “live3”. With it on, speech recognition is fixed to fish_audio:transcribe-stream-1. Current docs say that model also classifies overlaps as interruptions or acknowledgements.
How much does Live 3 cost?
The Live 3 page lists $0.035 per minute, down 50% from $0.07, until December 31, 2026. It says the plan includes Live 3. The pricing docs say the $0.07 list rate covers STT, TTS, basic LLMs, knowledge retrieval and tool calls.
Phone calls add $0.01 per minute. A number costs $1.20 per month. New accounts get a $5 trial balance.
How does Live 3 compare with other full-duplex voice models?
| Metrics | Fish Audio Live 3 | OpenAI GPT-Live-1 | Eleven Realtime v1 Max | Gemini 3.8 Live |
|---|---|---|---|---|
| Design | Full duplex, timing learned in the speech model | Full duplex, delegates reasoning to a backend model | Cascaded, tuned end to end | Live dialogue model, runs tools in the background |
| Languages | English only | Not disclosed | 90+ | 97 |
| Published timing | 186 ms minimum (short reactions) | Not disclosed in post text | 0.57 s avg time to first audio | Not disclosed |
| Turn-taking score | Not published | +30 pp Full Duplex Bench vs GPT-Realtime-2.1 (self-reported) | 94.1% Full-Duplex-Bench | Not disclosed |
| Reasoning / agent score | Not published | #1 on Tau3 with GPT-6 Astra (medium) | 98.8% Big Bench Audio; 87.1% tool tasks | Extended Thinking: 97.7% Big Bench Audio; 68.6% τ-Voice |
| Price | $0.035/min promo, $0.07 list (incl. STT, basic LLM, TTS) | $0.05/min voice layer + backend tokens | Session time + tokens | Not disclosed in launch post |
| Deployment | Agent Platform, API, on-prem | API, telephony | ElevenAgents, Realtime API | Gemini API, AI Studio, Gemini Enterprise (preview) |
ElevenLabs scores come from Artificial Analysis, as reported by ElevenLabs. OpenAI and Google figures are self-reported.
What it means
For developers building voice agents, Live 3 removes the need to manually tune silence thresholds to manage conversation flow. The system handles interruptions and confirmations automatically based on how people actually speak. However, the product currently supports English only and has not released independent benchmark scores to verify its claims against competitors.
