v2.1 — counting logic identical to v1; only the speech detector is upgraded to Silero VAD, running on-device. The app auto-detects the model's layout at load (works with Silero v4 or v5 ONNX files), self-corrects if the first configuration doesn't hear you, and falls back to the v1 energy gate — loudly logged — if the model can't run, so counting never silently stops. Assets are cached by a service worker (offline after first load). Use the “Voice activity” toggle to compare Silero vs the v1 energy detector in the same session.