Inherent chop-chop mic dojo

Turn on mic. Watch heads hit.

The bundled ONNX model loads with the page. Speak once and the mic signal, gate, and intent heads update in the same working view.

Model: loading Mic: idle
Loading bundled model Starting...
live oscilloscope Loading bundled model...

The model is loading. Recording unlocks automatically when it is ready.

Audio flow

Mic signal, mel frames, gate, and head hits stay together so changes are visible immediately.

Inference: idle
|_| mic raw PCM
->
~~~ mel 128 bins
->
{} gate waiting
->
{} heads intent hits

Donate labeled speech from the route

Open this drawer to reveal the per-head donation buttons and export locally saved JSONL.

0 saved
Per-head positive Interesting / no intent Hard speech negative Non-speech negative Confuser negative Ambiguous / review
Light validation gates
  • Record a short clip after the model is ready.

Record a clip, then press a head's donate button. Saved examples stay in this browser until exported.

Matched head

Scores are compared against the bundled model metadata thresholds.

No inference yet -- --

Runtime

This page ships with a same-origin ONNX model at assets/inherent.onnx. ONNX Runtime Web loads it with WebGPU when available and falls back to WASM.