Pipeline: mic → ASR (source) → translate (target) → speak (FIFO). One translation per VAD turn. Strong de-dupe.