[ 01 · INGEST ]
The recording becomes 2,048 numbers.
Audio is resampled to 32 kHz mono, turned into a log-mel spectrogram, and pushed through a frozen CNN14 trained on AudioSet. The network is never fine-tuned, on this site or any other. What leaves this stage is one vector per clip, and the waveform is never looked at again.
0 s1.5 s3 s
Each cell is one dimension, and the colour is how far this clip sits from what the site normally produces there, in standard deviations of the site's own background, to a limit of three. Dark is agreement, and most dimensions agree. Warm is above the site's usual level, cool is below. This grid, not the audio, is what every later stage operates on.
[ 02 · BASELINE ]
The site defines normal, in its own voice.
Every site sounds different, and a threshold tuned somewhere else does not transfer. So Canopy fits a Gaussian over the embeddings of this site's own unlabelled background: a mean vector and a covariance, shrunk ten percent toward its diagonal so it stays invertible. No labels are used. No one listens to the clips.
40 Hz500 Hz2 kHz8 kHz
The band is one standard deviation across clips. A wide band is a frequency this forest is inconsistent about, and the detector will forgive departures there.
85 background clips and the fitted two-sigma ellipse. Your selected clip is marked. In these raw coordinates a chainsaw lands inside the cloud, which is exactly the problem the next stage exists to solve.
[ 03 · SUBTRACT ]
Divide the forest out.
The baseline is removed by whitening. Subtract the site mean, then rescale every direction by how much this forest already varies along it. Directions the forest moves in freely get squashed; directions it never moves in get stretched. Afterwards the background is a ball of radius one, and how far a clip sits from the origin is exactly its Mahalanobis distance.
z = LT(x − μ) where Σ−1 = LLT and d = ‖z‖
Radius is the exact Mahalanobis distance. Bearing comes from a two-dimensional projection of the whitened space, so which way a point sits is approximate while how far out it sits is not. The ring is the alert line, and the scale is logarithmic because the loudest anomalies sit hundreds of times further out than the background does.
[ 04 · FLAG ]
The alert line is a budget, not a guess.
Ranger attention is the scarce resource, so the threshold is set by how many false alarms a site can absorb: ten percent of background. Crucially it is calibrated out of fold. The site's background is split five ways, each part scored against a density fitted without it, and the line drawn across the pooled result.
--
select a signal
- MAHALANOBIS DISTANCE
- --
- ALERT LINE
- --
- ANOMALY SCORE
- --
- BACKGROUND BUDGET
- --
WHY OUT OF FOLD MATTERS
Fit the Gaussian and read the threshold off the same clips and it comes back at 9.5. Score clips the fit never saw and the honest line is 251.4, twenty-six times further out. The embedding has 2,048 dimensions and the site has 85 recordings, so the fitted surface passes close to its own training points. An in-sample line would flag every unseen background clip and the ten percent budget would be a number on a slide rather than a number in the field.
[ 05 · ATTRIBUTE ]
Name it, or admit you cannot.
Only flagged clips reach this stage. The embedding is compared by cosine against one prototype per known class, built from verified positives elsewhere and containing no audio from this site. If nothing is similar enough, the mass goes to unknown and the alert still goes out. That is the point: a sound does not have to be recognised to be worth waking someone for.
CNN14 features are post-ReLU, so every embedding sits in the positive orthant and even unrelated classes score high against each other. The gate is the dashed line at 0.50; only the distance above it counts as evidence.
Reserved unknown mass is 1 − clip((max cosine − gate) / (1 − gate)), so a clip that resembles nothing known keeps its whole distribution on unknown rather than being forced into the nearest label.
[ 06 · MEASURED ]
What it actually does at this site.
The detector was fitted on Tembe background alone, then run over every threat clip available from other sites and public corpora, none of which contributed a single clip to the baseline. This table is the output of that run.
WHAT THIS DOES NOT SHOW