Fish Audio Launches Live 3 Full-Duplex Voice Model Responding in 186 ms

Executive Briefing
“Fish Audio has released the Live 3 full-duplex voice model, allowing virtual assistants to...”
Key Takeaways
- 1.Fish Audio's Live 3 model was trained on real conversations to learn conversational timing instead of relying on silence threshold rules.
- 2.The model's minimum response time reaches 186 milliseconds for short reactions like backchannels.
- 3.Live 3 currently supports English only and has not published independent benchmark scores.
Fish Audio has released the Live 3 full-duplex voice model, allowing virtual assistants to listen and speak simultaneously. The technology is designed to improve conversational timing and handle caller pauses.
Key Developments and Technical Context
Fish Audio's Live 3 model was trained on real conversations to learn conversational timing instead of relying on silence threshold rules.
The model's minimum response time reaches 186 milliseconds for short reactions like backchannels.
Live 3 currently supports English only and has not published independent benchmark scores.
Strategic Analysis and Industry Implications
Integrating turn-taking rules into the speech model rather than manual pipeline layers helps virtual assistants accurately distinguish real interruptions from backchannels, reducing mistimed responses.
The promotional pricing of $0.035 per minute alongside on-premises deployment options makes it easier for developers to test the service in practical applications.
Outlook and Key Factors to Watch
The timeline and scope for supporting languages other than English have not been specifically disclosed.
Comprehensive latency distributions and independent third-party benchmark evaluations have not yet appeared.







