v2 — counting logic identical to v1; only the speech detector is upgraded to Silero VAD, a small neural model running on-device (files cached by a service worker, so it works offline after the first load). It holds the burst through noise, quiet chanting and browser noise-suppression, and blocks music/taps/creaks from counting. If the model files are unavailable it falls back to the v1 energy gate automatically — use the “Voice activity” toggle to compare both detectors in the same session. Requires ort.min.js, ort-wasm-simd.wasm and silero-vad.onnx next to this page.