What the hydrophones actually heard

What was run

3 hydrophones, Bush Point, North San Juan Channel, Lab, over 6–6 September 2026, 1 calendar days and 3 node-days of audio: 8,572 thirty-second windows, 71 hours of audio. OrcaHello, a Southern Resident presence detector, flagged 1,001 of them as holding at least one positive segment, which is 11.7% of the audio. Every one was then classified for ecotype and measured for the resonance.

• Bush Point · Admiralty Inlet, off Whidbey Island · 48.03°N, 122.61°W · 24 h over 1 days
• North San Juan Channel · north San Juan Channel · 48.62°N, 123.09°W · 24 h over 1 days
• Lab · Haro Strait, west side of San Juan Island · 48.56°N, 123.17°W · 24 h over 1 days

Two counts exist and they are not the same. OrcaHello also returns a whole-window verdict, which fires on 762 windows; 239 further windows hold a positive segment without it. This page counts segments throughout, because a segment is what produces a clip to listen to.

Not included, because they are inactive in the feed registry rather than quietly missing:

• Orcasound MaST Center · node offline upstream; acoustic signal stopped 2026-07-12, segments continued until 2026-07-17
• Orcasound Point Robinson · hydrophone sediment-buried; the node streams but carries no usable signal

When each hydrophone heard something

Counts alone cannot tell an encounter from a fixed source. Per day they are different objects. A pod passing a node is loud for an hour and then gone; a machine is heard for a third of every day.

A named ecotype was claimed at Bush Point on 1 of its 1 recorded days, 15 hour(s) on each, every day it was listening; Lab on 1 of its 1 recorded days, 1 hour(s) on each, every day it was listening. Southern Residents pass Admiralty Inlet roughly once a month, so a claim on every recorded day at a single node describes a fixed source rather than an animal. The embedding test below reaches the same verdict by a different route.

The 4,816 Hz resonance

Frozen Perch 2.0 embeddings put Bush Point's saturated hours in a single tight non-biological cluster, 26.3 standard deviations tighter than true background, that no other hydrophone contributed a detection to. That said the audio was not whales. It did not say what it was.

It is a high-Q mechanical resonance: a fundamental at 4,816 Hz, 20.5 dB over true background, −3 dB width 62 Hz so Q = 78, carrying integer overtones at 2× and 3× and nothing at 0.5×. Across two days its centre holds to a standard deviation of 87 Hz, 1.8%. It is struck rather than running: 15.5 dB in windows the detector fired on against 1.4 dB in windows it did not. That is a metal object near the hydrophone, moved by water, ringing at its fundamental mode.

The spectrum below is not a single line at 4,816 Hz, and the claim does not need it to be. There is further raised structure between roughly 5 and 7 kHz, one part of it higher than either marked overtone. Struck plates and bars ring in inharmonic modes as well as integer ones, so extra peaks are what a metal object should look like rather than a problem for the reading. What carries the identification is the narrowness of the 4,816 Hz peak (Q = 78), the stability of its centre across days, its integer overtones, and the absent subharmonic, not its being the only feature in the band.

A Southern Resident call is tonal with a harmonic stack in roughly 1–6 kHz, and so is this, which is why spectral flatness, centroid, band ratios and pairwise spectrum similarity all failed to separate it. The sound is not un-whale-like in its spectrum. It simply never changes, and none of those measures were measuring change.

Before and after

The threshold is calibrated on this sample: 18 dB sits about 1.3 dB above the highest tone measured at the nodes where the object is absent. It is a filter for one object at one hydrophone, not a general de-noiser, and another site's contaminant will sit somewhere else in the spectrum.

Whether the numbers can be trusted

The ecotype classifier carries a measured false-positive floor on this audio: across 40 windows OrcaHello scored negative it produced 3 spurious Southern Resident calls and zero Bigg's. Every count here is tested against 7.5% for Southern Resident and against a Bigg's floor that 40 windows can only place below about 7.5%, using each hydrophone's own detections as the denominator.

• Bush Point · Southern Resident: 70 of 823 detections, 8.51%, against a floor of 7.5% (62 expected from error alone; p = 0.15).
• Lab · Southern Resident: 1 of 4 detections, 25.00%, against a floor of 7.5% (0 expected from error alone; p = 0.27).

Two limits on this test, both of which cut against reading it as proof. The floors are point estimates from small controls: Southern Resident is 3 of 40 windows, 95% CI 2.6%–19.9%, and Bigg's is 0 of 40, which is not a floor of zero but a floor no control this small can resolve below about 7.5%, so the Bigg's result is carried by the corroboration below and not by this test. And the test treats detections as independent draws, which a call bout is not: one encounter produces many correlated detections, which inflates any label that arrives in bouts. It is a screen, not a proof.

Does what survives hold up?

No embedding test was supplied, so the detections remaining after the filter have not been independently checked.

Does the independent record agree?

A classifier agreeing with itself is not evidence. These are the days the surviving detections concentrate on, set against what people on the water recorded the same day. Nothing below was used to produce the detections.

Nothing survived its own error rate, so there is nothing to check.

What survives, and what it sounds like

16 clips, drawn from detections that survived the filter. Cards the classifier named, at a site where that label clears its own false-positive floor, come first; the rest follow by OrcaHello's whole-window confidence, so a day where nothing clears the floor still offers audio to listen to. Every card states its own verdict, and no confident call means the classifier ran and named nothing. This is a listening sample and not a census. 16 of 16 clips are the full call bout rebuilt from the 30s source window.

Limitations

• OrcaHello gates everything downstream. 11.7% of this audio reached a classifier at all, and a Bigg's call has to sound Southern-Resident enough to trip a Southern-Resident detector.
• The rejection threshold is one number, calibrated on 1,001 tone measurements across 1 days at three nodes, with roughly 1.3 dB of margin. It is a filter for one object at one hydrophone.
• The false-positive floors are measured on windows OrcaHello scored negative. Those windows are quieter and less tonal than the windows a detection comes from, so the floors probably understate the rate that applies to the audio actually being labelled.
• The same ecotype classifier recovered only 5 of 20 killer whale recordings from the Watkins database, which is expected for a model specific to Salish Sea repertoires and means absence is weakly evidenced.
• These are the three live Orcasound nodes. MaST Center and Point Robinson are dead upstream and are listed above rather than omitted.
• The cards mix two orderings. Named calls that clear their site's floor come first, then the strongest remaining detections by OrcaHello confidence, which is uncalibrated and is a reason to listen first rather than evidence of a whale.