Fish Audio is launching Live 3 for teams building voice agents that handle live calls, including orders, bookings, and support. The model stays listening during its own speech, so a pause in a phone number, a brief "mm-hmm," or a change of mind halfway through a request can change what happens next. Fish Audio says those short reactions can start in as little as 186 milliseconds while a fuller answer is still being prepared. Live 3 puts conversational timing inside the speech model. The company trained it on real conversations so it learns when to acknowledge, wait, respond, or stop, instead of depending on a separate list of turn-taking rules. In Fish Audio's examples, it can answer between chunks of a number, keep going through a caller's "yeah," and stop when the caller actually cuts in with something like "wait, actually." English is supported at launch. Developers can run it on the Fish Agent platform and through the API, and enterprises can deploy it on-premises. Until December 31, 2026, Fish Audio lists Live 3 at $0.035 per minute, half the $0.07 rate, with phone calls billed as an add-on. The launch post also offers a free custom agent to companies that book a demo, and a free build for the first 1,000 people who reply with a use case. Rissa Cao, co-founder and CEO, announced it with co-founders Shijia Liao, chief scientist, and Jiahua Liu, chief technology officer. Fish Audio already sells text-to-speech, voice cloning, and speech-to-text. Live 3 is the release aimed at agents that have to hold the line while a person hesitates, revises, or talks at the same time.