Per-source EQ filters above 24 kHz for ultrasonic (bat) sources?

[SAdCodePrinter]

Hi, and thanks for BirdNET-Go and the bat support!

I'm not sure whether this is intended or whether I'm missing something, so I wanted to ask here first.

Setup

• BirdNET-Go nightly 20261014 (Docker, x86_64)
• Ultrasonic USB microphone (Vleermic 384k, UAC2) as its own audio source at 256 kHz, model bat
• A per-source equalizer on that source with a HighPass at 18 kHz works fine

What I tried
To suppress a few narrow, constant interference tones on that microphone (around 60, 64 and 72 kHz, most likely coming from its USB connection), I tried adding three BandReject filters to the source via PATCH /api/v2/settings/realtime:

equalizer:
  enabled: true
  filters:
    - { type: HighPass,   frequency: 18000, q: 0.707, passes: 4 }
    - { type: BandReject, frequency: 60000, width: 800, passes: 1 }
    - { type: BandReject, frequency: 64000, width: 800, passes: 1 }
    - { type: BandReject, frequency: 72000, width: 800, passes: 1 }

The request is rejected with 400 Invalid realtime settings. The same source without the three notches is accepted.

Looking at the code, I think this comes from MaxEQFrequency = 24000.0 in internal/conf/validate_audio.go ("Nyquist for 48 kHz"). If I read go-audio-dsp correctly, the filter itself would accept frequencies up to the Nyquist of the actual sample rate (128 kHz here), but I may well be missing something.

Why it matters for us
With these tones present, the bat model reports Pipistrellus pygmaeus about every 10–20 seconds whenever the source runs: roughly 2,500 detections per night, and even at midday. The ultrasonic validation marks almost all of them (about 99.9 %) as unlikely, which is great. We'd still like to try removing the tones before analysis. We're also working on the hardware side (cleaner USB power).

Question
Is the 24 kHz limit intentional for ultrasonic sources, or could it depend on the source's sample rate? Or is there another recommended way to filter in the ultrasonic range?

Happy to test a nightly build or share more details (spectra, config) if that helps. Thanks!

[tphakala]

That is just an oversight, not an deliberate limit. I can fix that for next release, for time being there are no nightly releases as code in main is not always stable to run at all.

Although what I have been experimenting with EQ is that most of the time it has only negative effects on classifier, so I am not very confident that EQ is right approach here. But please report back when you have had opportunity to try out fixed code. I'll try to do a next release soon.