I run a small independent ML monitoring project (oceansong.live) that
listens to the Orcasound feeds around the clock and scores them with OrcaHello. I think I've found a
likely physical cause for the Bush Point false positives behind #537, so I'm passing it along.
Short version: there's a repeating narrowband sound at Bush Point that looks and scores like a tonal
call, and to my ear it sounds like metal squeaking rather than an animal.
What the data looks like
Between 14 June and 28 July my system logged 2,375 orca detections at Bush Point, using my own gate
of OrcaHello >= 0.6, so the 90% change from #537 doesn't affect these counts. That's 28% of
everything flagged across nine hydrophones.
It's episodic rather than steady. Most days sit at 2 to 6 detections, but individual days spike hard:
2026-06-24 197
2026-07-04 189
2026-07-10 183
2026-07-21 222
2026-07-22 244
2026-07-23 317
2026-07-24 39 <- drops off
2026-07-28 4
The sound is a short squeak with several harmonics, repeating every second or two. Across 4,681
detected events the energy sits sharply at 4.4 to 5.2 kHz, with 59% between 4 and 5.5 kHz. Each
event has a fundamental plus three or four harmonics that all glide together in parallel.
Spectrogram from 4 July 13:34 UTC (per-frequency median removed), overview plus a zoom on one event:
https://oceansong.live/shared/bushpoint-squeak-20260704.png
The same shape repeats three times in 20 seconds.
Why I think it's mechanical
I listened to several of these clips. They sound like metal parts squeaking against each other. I
wouldn't claim certainty from 16 kHz audio, and one clip I'd call genuinely ambiguous, but every
Bush Point clip I've checked by ear so far has come out that way, across separate episodes on 4
July, 21 July and 29 July.
I don't think this reflects badly on the model. A squeaking metal fitting really does produce the
same gliding harmonic stack that a tonal call does, so there may be no spectral feature that
separates them. A small CNN I trained myself fires on it too, and equally can't tell the difference.
Why the threshold change may only go part way
Checking my own detections against the 90% threshold from #537:
| threshold |
Bush Point detections surviving |
| 0.60 |
2,375 (100%) |
| 0.85 |
1,224 (52%) |
| 0.90 |
978 (41%) |
| 0.95 |
632 (27%) |
More to the point, of the clips I picked out as squeak-like, the OrcaHello confidences were 0.64,
0.69, 0.77, 0.95 and 0.98. Two of five would still pass at 90%. Across a wider sample the
correlation between OrcaHello confidence and whether a clip actually contains call structure came
out at +0.04, so confidence isn't separating signal from this artifact.
If the source is physical, finding and quietening it (or moving the hydrophone away from it) would
probably do more than threshold tuning.
What it might be
I genuinely don't know. The episodic pattern would fit a moored vessel with metal fittings working
in the swell, or a mooring chain or shackle, or something on the mount itself. The drop after 23
July might mean whatever it was moved or stopped. Someone who knows the site will have much better
intuition than I do here.
Happy to share the clips, the raw timestamps, or the per-day counts if any of that is useful.
I run a small independent ML monitoring project (oceansong.live) that
listens to the Orcasound feeds around the clock and scores them with OrcaHello. I think I've found a
likely physical cause for the Bush Point false positives behind #537, so I'm passing it along.
Short version: there's a repeating narrowband sound at Bush Point that looks and scores like a tonal
call, and to my ear it sounds like metal squeaking rather than an animal.
What the data looks like
Between 14 June and 28 July my system logged 2,375 orca detections at Bush Point, using my own gate
of OrcaHello >= 0.6, so the 90% change from #537 doesn't affect these counts. That's 28% of
everything flagged across nine hydrophones.
It's episodic rather than steady. Most days sit at 2 to 6 detections, but individual days spike hard:
The sound is a short squeak with several harmonics, repeating every second or two. Across 4,681
detected events the energy sits sharply at 4.4 to 5.2 kHz, with 59% between 4 and 5.5 kHz. Each
event has a fundamental plus three or four harmonics that all glide together in parallel.
Spectrogram from 4 July 13:34 UTC (per-frequency median removed), overview plus a zoom on one event:
https://oceansong.live/shared/bushpoint-squeak-20260704.png
The same shape repeats three times in 20 seconds.
Why I think it's mechanical
I listened to several of these clips. They sound like metal parts squeaking against each other. I
wouldn't claim certainty from 16 kHz audio, and one clip I'd call genuinely ambiguous, but every
Bush Point clip I've checked by ear so far has come out that way, across separate episodes on 4
July, 21 July and 29 July.
I don't think this reflects badly on the model. A squeaking metal fitting really does produce the
same gliding harmonic stack that a tonal call does, so there may be no spectral feature that
separates them. A small CNN I trained myself fires on it too, and equally can't tell the difference.
Why the threshold change may only go part way
Checking my own detections against the 90% threshold from #537:
More to the point, of the clips I picked out as squeak-like, the OrcaHello confidences were 0.64,
0.69, 0.77, 0.95 and 0.98. Two of five would still pass at 90%. Across a wider sample the
correlation between OrcaHello confidence and whether a clip actually contains call structure came
out at +0.04, so confidence isn't separating signal from this artifact.
If the source is physical, finding and quietening it (or moving the hydrophone away from it) would
probably do more than threshold tuning.
What it might be
I genuinely don't know. The episodic pattern would fit a moored vessel with metal fittings working
in the swell, or a mooring chain or shackle, or something on the mount itself. The drop after 23
July might mean whatever it was moved or stopped. Someone who knows the site will have much better
intuition than I do here.
Happy to share the clips, the raw timestamps, or the per-day counts if any of that is useful.