The pre-wearable era
Classical field bioacoustics had two sources of data: stationary microphones (parabolic dishes, distance-attenuated, recording a wide acoustic scene with no per-individual attribution) and synchronized observation logs (a researcher writing down what individual crows were doing while audio recorded in parallel). The mismatch was the bottleneck. A stationary microphone twenty meters from a foraging crow captured the call but not the eye contact, the head position, the partner's response. A researcher with a clipboard captured the behavior but lost the acoustic detail to attenuation and ambient noise. The two streams could be joined statistically — there's a long literature on this — but the joining was always lossy, and the loss was specifically at the resolution that mattered for behavioral interpretation.
Vlad Demartsev, Ariana Strandburg-Peshkin, and colleagues fitted wild carrion crows (Corvus corone) with lightweight wearable bioacoustic loggers — devices small enough to ride on the bird without altering behavior, with onboard storage and a clock synchronized to a base station.
What Demartsev et al. did
Vlad Demartsev[1], Ariana Strandburg-Peshkin, and colleagues fitted wild carrion crows (Corvus corone) with lightweight wearable bioacoustic loggers — devices small enough to ride on the bird without altering behavior, with onboard storage and a clock synchronized to a base station. The loggers recorded the bird's own audio at close range, eliminating attenuation, ambient noise from other sources, and the per-individual attribution problem. Synchronized behavioral observation from human researchers continued in parallel, but the audio side of the equation was suddenly per-individual and high-fidelity. Over the study period, the team accumulated 127,000+ vocalizations across diverse behavioral contexts, all attributable to specific known individuals.
Why the dataset matters
Three things change when you have 127,000+ per-individual recordings with synchronized behavior. First, sample sizes go from 'one researcher's PhD' to 'pre-training-scale.' You can train -based context classifiers on the data and get meaningful results. Second, individual-level questions become tractable: how do this bird's calls change across years, across mates, across breeding versus non-breeding seasons? Third, sequence-structure questions become tractable: is a call's meaning a function of the calls before and after it in time? The fifty-year debate about whether crow vocalizations have anything like compositional structure was constrained by the data, and the data has just gotten much better.
What the paper found
Demartsev[1] et al.'s 2026 preprint mapped the carrion crow repertoire with Voxaboxen (an open-source bioacoustic segmentation tool) and projection over BEATs embeddings, then trained a behavioral-context classifier on the synchronized data. The result was a repertoire map with cluster-level behavioral probabilities — territorial, mobbing, affiliative, recruitment, individual-specific contact — at finer resolution than any prior bird study. Crucially, the same call type took on different cluster centroids depending on social context: a call given alone differed acoustically from the same call type given to a mate. The contextual modulation was small but consistent across individuals. That's the cleanest evidence to date that wild corvid vocalizations carry social-context information beyond their categorical type.
Does it generalize to American crows
Yes and no. The carrion crow (Corvus corone) and the American crow (Corvus brachyrhynchos) are sister species with similar social systems, similar vocal repertoires, and similar cognitive baselines. Findings about social-context modulation should transfer in spirit, though the specific acoustic dimensions that modulate may differ. The bigger constraint is methodological: no comparable wearable-logger study has been published for American crows as of mid-2026. CrowLingo's atlas is built on stationary-microphone CC-licensed recordings, which means our cluster-level behavioral probabilities are downstream of older literature, not synchronized observation. We're transparent about this. The eventual American crow Demartsev[1]-equivalent study — if and when it happens — will refine our behavioral-context probabilities meaningfully.
Why the methodology is the news
It's tempting to focus on the findings — contextual modulation of call meaning is interesting — but the methodology is the bigger story. Wearable bioacoustic loggers will become a standard tool in corvid research over the 2026-2030 horizon. The Demartsev[1] paper isn't a one-off; it's the proof of concept for a research program that other labs (notably Cornell, Marzluff's UW group, and the Wright lab at New Mexico) are already adopting for their species of focus. Five years from now, the question won't be whether wild bird communication can be studied at per-individual resolution. It will be: which species has been instrumented this way, and what did we find.
What this means for AI bioacoustic models
Pre-2026 bioacoustic models were trained on stationary-microphone audio, which means they learned representations of acoustically-mixed scenes — multiple birds at varying distances, overlapping calls, background noise. Wearable-logger audio is fundamentally different: single-bird, close-range, low noise. Models trained on or fine-tuned for this data will have cleaner geometries. Whether the geometries reproduce the stationary-microphone-trained models' findings or diverge from them in informative ways is one of the empirical questions the field will answer over the next two years.