Fish Audio has introduced Live 3, a voice model for agents that listens while it speaks. The company says the model can acknowledge a caller, wait for them to finish, or stop when they change direction. Its shortest reported response time is 186 milliseconds.
That approach is meant to handle the uneven rhythm of ordinary calls. Someone might pause while recalling a phone number, say “yeah” to show they are listening, or revise an order before finishing the sentence. Fish Audio says Live 3 can respond with a brief “mm-hmm” or “okay” without taking over, and keep speaking through a simple acknowledgment.
Timing is part of the model
Fish Audio argues that a quick reply is not always a well-timed reply. Many voice agents coordinate speech recognition, reasoning, speech generation and turn-taking across separate parts of a system. Live 3 instead brings conversational timing into the speech model, the company says.
The model was trained on human conversations to learn when to acknowledge, wait, respond or stop, according to Fish Audio co-founder and CEO Rissa Cao. The company’s examples include a caller giving a number in chunks and changing a pizza order midway through. These are demonstrations from Fish Audio, not independent tests of how Live 3 performs across different calls or noisy environments.
Where can people use Live 3?
Fish Audio says Live 3 is available through its Agent Platform and API, with on-premises deployment available for enterprises. It supports English at launch. The company says customers provide the instructions, tools and workflows for their agents; Live 3 supplies the speech and conversational timing.
Why it matters
For businesses building phone agents for bookings, orders or support, Live 3 is designed to let callers correct themselves or interrupt without restarting the exchange. Fish Audio has not provided independent performance results, so how reliably it handles real calls remains unclear.
