Live 3
A full-duplex speech model for voice agents introduced by Fish Audio. It supports listening and speaking simultaneously, can distinguish backchannel utterances from genuine interruptions, and provides low-latency response capability.
Timeline (last 90 days)
Fish Audio released Live 3, a full-duplex speech model aimed at voice agents.
Live 3 can listen continuously while speaking, i.e., it has a full-duplex architecture.
Live 3's response latency is reduced to 186 milliseconds.
Live 3 supports only English at launch.
Live 3 claims to distinguish meaningless backchannel sounds from genuine interruptions, treating backchannel sounds as permission to keep speaking and genuine interruptions as a signal to yield the turn.
According to a report by MarkTechPost, Live 3 is designed for voice agent scenarios, emphasizing real-time responsiveness during conversation rather than merely pursuing the naturalness of speech synthesis.
The information Fish Audio disclosed this time focuses on capabilities and latency; details such as model size, training data, deployment method, and pricing were not elaborated in the original report.
All Facts (7)
Live 3 supports only English at launch.
50%UnverifiedLive 3 can listen continuously while speaking, i.e., it has a full-duplex architecture.
50%UnverifiedLive 3's response latency is reduced to 186 milliseconds.
50%UnverifiedThe information Fish Audio disclosed this time focuses on capabilities and latency; details such as model size, training data, deployment method, and pricing were not elaborated in the original report.
50%UnverifiedLive 3 claims to distinguish meaningless backchannel sounds from genuine interruptions, treating backchannel sounds as permission to keep speaking and genuine interruptions as a signal to yield the turn.
50%UnverifiedAccording to a report by MarkTechPost, Live 3 is designed for voice agent scenarios, emphasizing real-time responsiveness during conversation rather than merely pursuing the naturalness of speech synthesis.
50%UnverifiedFish Audio released Live 3, a full-duplex speech model aimed at voice agents.
50%