Each audio piece is stamped twice: when it
arrives from the cloud model (recv) and when this
server
sends it to the phone bridge (send). A gap of
250 ms+ counts as a stall.
Alibaba cloud --recv--> THIS server --send--> bridge -----> caller
cloud — the gap was already there when audio arrived → Alibaba generation paused.
this-server — arrived smoothly but sent late → this server was slow to relay (e.g. event-loop block).
mixed — both show gaps.
clean — no stall here; if the caller still heard one, it was the network to the bridge (check the bridge's own log).