Fish Audio shipped Live 3 on November 10, 2026, introducing a full-duplex voice model designed specifically for agents. The company claims the model can decide to emit a short reaction in as little as 186 ms, learning turn-taking directly from real conversations rather than relying on traditional VAD thresholds. While this is a significant leap for cloud-based voice AI, developers building local-first applications must recognize that this latency metric only applies once clean PCM data reaches the model. The real bottleneck for many desktop applications remains the Windows audio capture pipeline.
The Illusion of Cloud Latency
Hosted duplex models like Live 3 cannot see the local audio environment. They do not have access to private microphone tracks that bypass the call, nor can they generate translation overlays invisible to screen sharing. For developers needing real-time speech-to-text during technical interviews or support calls, sending audio off-box introduces network round-trip time that the 186 ms model latency does not account for. The practical constraint is not the AI's decision speed, but the integrity and timing of the audio stream delivered to it.
Dual-Track WASAPI Is the Practical Floor
To maintain speaker identity and enable independent processing, developers must use dual-track capture. This involves separating the microphone input from the loopback render device using the Windows Audio Session API. The Windows Audio Engine (audiodg.exe) routinely adds 10โ40 ms of delay due to mixing and vendor APOs, a cost that persists regardless of the AI model's speed. Even with optimized capture clients, buffer sizes of 10โ40 ms are common to prevent drops during DPC spikes, meaning end-to-end partials often land in the 80โ150 ms range before any network transmission occurs.
Key Takeaways
- Live 3's 186 ms reaction time is valid only for clean PCM input, ignoring local capture and network delays. Dual-track WASAPI capture is required for accurate speaker separation in real-time translation and transcription apps. Windows Audio Engine processing and buffer constraints add significant latency that cloud model benchmarks omit. Local processing with Whisper.cpp or BYOK endpoints may offer better control over latency than pure cloud solutions. Exclusive mode offers lower latency but breaks compatibility with other apps, making shared mode the pragmatic default.
The Bottom Line
Don't let the 186 ms marketing metric fool you: for local-first apps, the Windows audio stack is the true latency governor. Optimize your dual-track capture pipeline before worrying about cloud model response times.