Abnormal noises a seasoned ear can hear —
now machines can hear them too.

A machine can't compare two sounds directly, so it first measures a 4-second recording from five different angles, then joins the readings end to end into one "fingerprint". Whether two units are alike comes down to how close those ~400 numbers are.

Every reading carries physical meaning — which band holds the energy, whether it sounds rough, whether the bands rise and fall together, the millisecond-level repetition in the waveform, and how far it drifts from this product's own normal. If a decision goes wrong, you can trace it back to the exact reading responsible.

The 398 readings from one audio clip, laid end to end in five groups. Color depth = that reading's rank within its group.

The perspective that works best differs from one inspection point to the next. At four points in one auto-parts plant, a perceptual metric was the most accurate; at most points in an audio-equipment plant and a bearing plant, the most basic "band energy distribution" was more accurate instead.

So rather than betting on one universal perspective, we measure all five and combine them — at 11 of 12 real inspection points, the combined result was no worse than any single perspective.

12 real inspection points × 5 perspectives + a combined column. 5-fold out-of-sample cross-validation.

Apply one fixed threshold across 12 points and the share of good units wrongly blocked ranges from 0.9% to 14.4% — not because the model is inaccurate, but because every production line has its own noise floor, operating conditions, and microphone placement.

The fix is to convert each score into "which percentile it sits at among this point's good units", then set the threshold against a unified Error budget. After conversion, the same 12 points all land within 10.0%–10.5% — a rule set on one line can be handed straight to another.

Left: one fixed threshold. Right: threshold set from a 10% Error budget after conversion to per-point percentiles. Both plots share the same axes and scale.

The first two weeks after go-live are the least stable: samples are few and the threshold is not yet dialed in, so detection runs low and wrong-blocking runs high. The system keeps calibrating on the labels it accumulates day by day and converges into the usable range in about a month.

Once stable, the typical range is a detection rate around 90%, with the wrong-block rate back near its 10% budget. You don't need long to get started — on the measured learning curve, about 10 nonconforming samples already reach close to 80% detection, and every further batch pushes it higher.

Real labeled days at one inspection point in an auto-parts plant, where each day's units were judged using only the knowledge available up to the day before. This chart shows only the run-in to convergence; the curve is a rolling average, reflecting the trend rather than any single day's result.