Turn-taking takes center stage in Fish Audio’s Live 3 voice model

The Live 3 voice model from Fish Audio listens while speaking and adapts to pauses. It is available in English.

A “yeah” during an explanation can simply signal that someone is following along. A “wait” asks for the floor back. Distinguishing between these intentions is central to Live 3, Fish Audio’s full-duplex voice model, which keeps listening while it speaks. The company is targeting support, booking, and order-taking agents, where a poorly timed response can cut a request in half.

The examples presented focus on these everyday exchanges. While a caller reads out a phone number in groups of digits, the agent can offer a brief acknowledgment and wait for the rest. When an order changes midway through a sentence, it is supposed to let the caller finish the revisions before responding. These behaviors are illustrated in three excerpts the company describes as recordings from test calls, preserving their original pauses and overlapping speech. Learning when to speak

According to the company, turn-taking is integrated into the speech model and learned from human conversations. Training is intended to teach it when to wait, acknowledge, respond, or stop speaking. A separate reasoning layer handles the agent’s instructions and tools: managing the spoken exchange remains distinct from processing the business request.

The company reports that the model can decide to respond in as little as 186 milliseconds. This figure concerns the decision to speak and accompanies examples of short reactions, such as “mm-hmm” or a quick confirmation. It does not measure the time needed to produce a complete answer after looking up a record or performing an action. The post provides neither a detailed measurement protocol nor a distribution of response times across calls.

Access is available through the Fish Agent Platform and API, with on-premises deployment offered to enterprises. At launch, the model speaks English. References to more than 30 languages and millions of voices on the commercial page describe the platform’s broader catalog; they do not establish multilingual support for Live 3. Connecting calls to documents and business tools

To turn this voice layer into a service agent, the agent-building platform, currently in public beta, brings together instructions, voices, knowledge sources, and tools. Businesses can supply their information, connect their APIs, and integrate an agent into a website or phone number. Booking appointments, checking orders, or issuing refunds depends on those connections and the actions configured.

The advertised rate, which includes Live 3, is $0.035 per minute during a promotion scheduled to run through December 31, 2026, compared with a standard rate of $0.07. Phone calls add $0.01 per minute, and a dedicated number costs $1.20 per month. Using a premium reasoning model incurs additional token-based charges. The headline price therefore does not, on its own, represent the full cost of every call configuration.

Before publishing, agents can be tested using scripted conversations and tool calls. After deployment, transcripts, recordings, and action histories support conversation monitoring. The documentation also specifies that each published release freezes a numbered version: a configuration being edited remains separate from the one used for production calls.