Fish Audio has introduced Live 3, a voice model designed to keep listening while an agent is speaking. The company says it can distinguish an interruption from a short acknowledgment, wait while a person finishes a thought and respond without forcing the caller through rigid conversational turns.
The November 10 announcement describes an English-language model available on the Fish Agent Platform and through its API. Fish Audio also says enterprises can deploy it on premises; that statement alone does not establish that model weights are publicly downloadable. This article was published November 13.
Why conversational timing matters
A conventional voice assistant may connect speech recognition, an instruction-following model, speech synthesis and rules for deciding when to speak. Those boundaries can make a quick acknowledgment interrupt an answer or cause an agent to respond while someone is still giving a phone number. Fish Audio says Live 3 incorporates conversational timing into the speech model while a separate reasoning layer handles the agent’s instructions and tools.
The company says the model was trained on real conversations to learn when to acknowledge, wait, respond or stop. Its examples include acknowledging pieces of a number without taking over, and yielding when a caller changes an order. These are product demonstrations and company descriptions, not independent evidence that every interruption will be handled correctly.
Latency, access and tradeoffs
Fish Audio says Live 3 can decide to respond in as little as 186 milliseconds. That is the company’s reported minimum for the behavior it describes, not a guaranteed end-to-end response time for an application. Network conditions, the reasoning model, audio processing and deployment choices can all affect the delay users experience.
For developers, an agent that continues to listen raises practical questions about microphone indicators, consent, retention and how recordings or transcripts are handled. Check the provider’s current documentation, terms and pricing, and test the actual deployment with background noise and varied speakers before relying on it.
What remains unverified
Fish Audio provides examples of test calls in its announcement, but AiBrainWorX has not independently tested Live 3, measured latency or assessed its accuracy and privacy behavior. The company identifies English as the language available at launch; readers should check for later changes rather than assume multilingual support.
Original reporting based on publisher announcement; no independent product test.
