What the study actually measured
Mates[1] and colleagues analyzed 1,674 selected caws from 18 marked wild American crows in Cayuga Heights, New York. The recordings came from June through August 2006. Calls were chosen for high signal-to-noise ratio, a visible identified caller, and no overlap with other birds. A subset of 305 calls had behavioral contexts identified by observation. Those selection conditions matter: the results do not describe arbitrary noisy recordings or every crow call type.
The authors measured acoustic features, reduced them with principal component analysis, and used linear discriminant analysis to classify callers.
Individual identity is a limited result
The authors measured acoustic features, reduced them with principal component analysis, and used linear discriminant analysis to classify callers. Their Results section reports 35.36% correct identity classification on the fitted data and 24.24% with cross-validation, significantly above chance. The latter is the more useful number when discussing performance beyond fitting the data. It is a result among the study’s known birds, not near-perfect identification, recognition of an unknown crow, or evidence that a phone app can identify your neighborhood birds.
A voice difference is not a regional dialect
Individual variation, group differences, learned conventions, and a listener’s response are different questions. This study’s recordings cover one area and one summer. It does not compare geographically separated populations across years, establish stable family-group acoustic centroids, or test whether young crows learn a local dialect. Its evidence cannot decide whether a regional difference is cultural, genetic, environmental, or some combination. Those possibilities require their own study designs and sources.
What modern AI would still need to show
This paper did not evaluate , , , or a modern audio pipeline. A new model’s ability to produce vectors does not establish caller recognition or dialect detection. A useful evaluation would define the task, identify the birds independently, report the recording conditions and candidate population, and test performance on appropriately separated recordings. Repeated calls from the same recording session can make a test less demanding than deployment in a new setting. Those are evaluation requirements, not results obtained by CrowLingo.
What would strengthen a dialect claim
A regional-dialect claim needs recordings from the populations being compared, comparable call contexts, and a method that separates geography from differences in callers, recording equipment and habitat. A claim of cultural transmission needs evidence about learning; a claim that differences matter to listeners needs behavioral evidence. This is a description of the evidence needed, not a finding that the wider literature has no such work. This article reviews the cited Mates[1] study, not every publication on corvid vocal learning.
Listen with the right expectations
You can compare archival recordings and notice pitch, timing and texture without assigning an individual identity or a regional meaning. CrowLingo’s Atlas has illustrative coordinates and editorial groups; it does not run a caller-identification or dialect classifier. Use headphones and observe wild birds without broadcasting calls to them. The primary study is linked in the references below so you can check its methods and limitations directly.