Fish Audio Launches Live 3 Voice Model

Live 3 is a full-duplex voice model that keeps listening while it speaks, aimed at real-time voice agents.

Updated on Nov. 11, 2026 in Artificial Intelligence

Isometric editorial illustration of two matte black microphones facing each other across a pale table, linked by a single coiled cable.
Fish Audio released Live 3, a full-duplex voice model that keeps listening while it speaks, moving real-time turn-taking into the speech model itself. AI Illustration. Upload story photo >

Fish Audio released Live 3, a full-duplex voice model for real-time voice agents, and made it available on its Agent Platform, through the Fish Audio API, and on-premises for enterprises. The model keeps listening while it speaks and moves conversational turn-taking into the speech model itself.

Why it matters

Speed and timing are separate problems, and an agent can reply quickly while still speaking at the wrong moment. Live 3 is designed to address both by learning when to acknowledge, wait, respond or stop from training on real human conversations.

Fish Audio's company-published specifications list a decision time of 186 ms for short reactions and a library of over 2,000,000 voices including cloned voices. Fish Audio has not published a latency distribution, and no independent Artificial Analysis score had been reported as of November 11.

The players

Fish Audio

Fish Audio is a voice AI company that offers text-to-speech and voice cloning products, including the Fish Audio API and an Agent Platform for building voice agents.

The details

Live 3 runs as a full-duplex model, so it continues listening during its own speech rather than relying on separate pipeline rules to manage turns. A separate reasoning layer holds the agent's instructions and tools, while the speech model handles timing, such as saying "okay, 415" between chunks of a phone number and then waiting.

Timeline

  1. November 10, 2026: Fish Audio releases Live 3.

  2. December 31, 2026: The introductory price of $0.035 per minute ends.

The Tech Race

Live 3 places Fish Audio in the full-duplex voice model category, where turn-taking is handled inside the speech model instead of separate pipeline rules. This moves the design away from turn-based voice pipelines that rely on rule-based interruption logic.

Developers building voice agents can now license Live 3 through Fish Audio's API or on-premises deployment, with an introductory rate of $0.035 per minute until December 31, 2026. New accounts receive a $5 trial balance, and a phone number costs $1.20 per month.

The takeaway

For teams building voice agents, timing control is now a model-level capability, so evaluating conversational timing, not just response speed, is worth prioritizing. Fish Audio's published figures cover speed and pricing, but independent latency and quality benchmarks are still pending.

Further reading

Explore more Artificial Intelligence coverage to see how voice models are changing real-time conversational systems.

More information

Try the voice model on the Live 3 product trial page.

Source note: This article includes information reported by MarkTechPost.