Start with the sample
The authors recorded marked wild American crows in Cayuga Heights, New York, during June–August 2006. They selected 1,674 tonal, harmonically rich caws from 18 birds: 11 male and seven female. Selection required high signal-to-noise ratio, a visible caller identified from its marks, and no overlap with other birds. All sampled birds were at least a year old. This is a particular study sample, not a benchmark covering all ages, populations, call types or recording conditions.
The study measured 23 acoustic parameters spanning pitch, power-envelope and spectral properties and used principal component analysis to summarize them.
How the analysis worked
The study measured 23 acoustic parameters spanning pitch, power-envelope and spectral properties and used principal component analysis to summarize them. Mixed-effects models examined relationships with sex, age and observed behavioral context. Linear discriminant analysis assessed classification by identity and sex using selected components. The authors deliberately avoided dividing their calls into discrete call types because they observed continuous variation. Describing this as an AI discovery of nine call categories would misrepresent the method.
Read the identity result with its denominator
For caller identity, the Results section reports 35.36% correct classification on the fitted data and 24.24% with cross-validation. The authors report that the latter is significantly above chance among the 18 known birds. Both the denominator and the validation method matter. Above-chance classification shows that the measured calls contain some information about identity; it does not mean every bird is uniquely recognizable, that performance generalizes to an unseen population, or that the birds themselves use the same cues.
Context and sex are separate findings
Behavioral context was recorded for a subset of 305 calls, in food recruitment, territorial counter-cawing, beg rebuff and alarm situations. The analysis found context-associated acoustic differences and sex-related variation. These are associations within the sampled observations, not a decoder of intent from any half-second of audio. The abstract describes sex and identity prediction as moderately accurate. The analysis found no significant main effect of age, but it did report a sex-by-age interaction for one component. That is not evidence that age never matters.
What the paper did not test
The recordings do not provide a multi-region, multi-year dialect comparison. The study did not test cultural transmission of local vocal conventions, listener comprehension, lexical meaning or grammar. It also did not evaluate , , or a self-supervised pipeline. Claims that every modern model reproduces its results need separate, named evaluations. Acoustic classification and evidence about what a signal means to a crow should not be collapsed into one claim.
Why it remains useful to a learner
The paper offers a concrete example of linking measured sound features to independently observed callers and contexts, with uncertainty and selection limits stated. That makes it useful for learning how to assess a bioacoustic claim. Read the sample and Methods before repeating a headline accuracy figure. When comparing a newer study, check whether the species, task, population and validation design actually match. CrowLingo has not replicated this experiment; its Atlas is an illustration for listening, not an individual-recognition result.
Correction and source access
This article was corrected on November 11, 2026. The earlier version overstated recognition accuracy, year-to-year stability and replication by modern models. Those claims were not supported by the primary paper inspected for this revision. The linked full text is the basis for the numbers and methods described here. This source check is not a systematic review or an independent ornithological peer review.