01 / TURN STATEThe face follows the conversation
Turn state, not audio amplitude, decides where it looks and when it shows processing. Mouth articulation is derived from the actual output audio, so it cannot drift from speech.
0.65 ms median
0.83 ms p95Expression inference on CPU. 25 ms causal analysis window, 10 ms recurrent updates.
02 / PLAYBACK CLOCKAudio is authoritative
Expression frames travel immediately before the PCM packet they describe, addressed in 16 kHz samples. If rendering stalls, stale poses are dropped instead of replayed late.
~155 msSource canvas to decoded WebRTC at 30 fps.
03 / ASYNC WORKTools do not stall the turn
The model dispatches work and keeps talking. Results arrive when the conversation is next idle; the video above shows the longer call returning while the session continues.
22.1 s and 9.7 sTwo concurrent think calls in the recorded demo session. Neither blocked a conversational turn.
04 / SPECULATIONSilence is spent generating
Presence starts generating at a provisional endpoint while it keeps listening. Resumed speech discards the work. Confirmed silence releases audio already made.
1.8–2.1 s p50
~1.2 s speculativeWarm full turn, microphone speech to first audio. A native-live baseline was faster on the same fixture: 1.054 s versus 1.955 s.