Fish audio launches live 3, a new full-duplex voice model capable of sub-second reactions

Fish Audio has introduced Live 3, a full-duplex voice model designed for real-time voice agents, allowing the system to listen and speak concurrently. This architecture shifts conversational turn-taking logic from rigid pipeline rules directly into the speech model, which was trained on human conversations to manage acknowledgments and responses. Live 3 can react in as little as 186 ms for brief acknowledgments like 'okay' or 'mm-hmm,' with complete answers being prepared in parallel. While the company has not released a latency distribution, third-party tracker OrcaRouter reported no Artificial Analysis score for Live 3 as of November 11, 2026. The model is accessible via the Fish Agent Platform, the Fish Audio API, and on-premises deployments. The platform supports over 2,000,000 library voices, including cloned options. Pricing for the service is set at $0.035 per minute, a 50% reduction from the previous $0.07 rate, valid until December 31, 2026. Phone calls incur an additional $0.01 per minute, and a number costs $1.20 monthly.