Fish Audio releases Live 3, a full-duplex voice model that reacts in as little as 186 ms

Fish Audio released Live 3, a full-duplex voice model for voice agents, which can react in as little as 186 ms, according to MarkTechPost's report on the launch.

What Live 3 does

Live 3 is built to listen while it speaks. It distinguishes short backchannel sounds such as "mm-hmm" from real interruptions, a distinction voice agents need to make when a user talks over them.

Key details

• Model type: full-duplex voice model for voice agents
• Reaction time: as little as 186 ms
• Language support: English only at launch
• Pricing: not yet disclosed
• Context window, parameter count and benchmark scores: not yet disclosed

What this means

Voice agents that handle audio in one direction at a time can struggle to tell a listener's acknowledgement from an attempt to take the floor. Live 3 targets that problem by listening during its own speech and separating "mm-hmm" from real interruptions. The 186 ms figure is a best-case reaction time ("as little as"), so teams evaluating the model should test latency under their own conditions. English-only support at launch limits deployments to English-language use for now.