audio-verification

Audio verification

Code that compiles is not sound that works. You cannot hear anything, so an unmeasured audio change is an unverified one. Reading the graph, passing typecheck, and "the node is connected" prove none of: it makes sound at all, it isn't a burst of noise, it doesn't click on attack, it isn't clipping, it isn't inaudibly quiet.

The good news: audio is just numbers. Render the graph offline and every perceptual complaint becomes a measurement.

The rule

Before saying a sound change is done, fixed, or good:

1. Render it offline — OfflineAudioContext gives you the exact samples, deterministically, far faster than real time.
2. Measure it — peak, spectral flatness, onset, tail (see below).
3. A/B against the version you changed — almost every judgement here is relative, not absolute.
4. If it doesn't match intent, iterate. If you genuinely can't measure it, say so plainly instead of implying you checked.

Running the probe

scripts/audio-probe.js is a classic script (no imports) so it pastes straight into page.evaluate(), a browser-MCP javascript_tool call, or a devtools console. Evaluating it defines globalThis.AudioProbe.

// build(ctx, destination) wires and starts your nodes — anything you'd build live
const m = await AudioProbe.measure((ctx, dest) => {
  const osc = ctx.createOscillator();
  const g = ctx.createGain();
  g.gain.setValueAtTime(0, 0);
  g.gain.linearRampToValueAtTime(0.3, 0.02);
  g.gain.exponentialRampToValueAtTime(0.0004, 0.4);
  osc.connect(g).connect(dest);
  osc.start(0); osc.stop(0.5);
}, { seconds: 0.6 });

// compare variants under identical conditions
const table = await AudioProbe.compare({ before: buildOld, after: buildNew }, { seconds: 0.6 });

Any page will do — about:blank is enough, you don't need the real app to measure a graph. To measure the real code path, either import the module under test, or copy its graph verbatim into build (and say which you did; a hand-copied graph can drift from the source).