Continuous variation was already measurable
In their study of 18 marked wild American crows, Mates and colleagues used 23 acoustic measurements, principal component analysis and statistical models. They reported continuous variation and did not divide the caws into discrete call types. Their results found associations with caller identity, sex and behavioral context within the study's sample and analysis.
That study is an example of automated measurement and continuous analysis before current audio foundation models. Its findings do not establish a general-purpose crow identifier or a translation of call meaning. Read the Mates et al. primary paper and our study guide for the sampling and evaluation limits.
Learned features still need a defined task
Animal language processing includes several training approaches. In Audio-MAE, masked spectrogram reconstruction supplies the pretraining objective; the paper then uses labeled data for supervised fine-tuning in its classification experiments. Those audio and speech benchmark results are specific to the tasks, datasets and comparisons reported.
Other systems use different objectives and supervision. A claim that one representation performs better needs a named model, a comparator, an evaluation set and a metric. CrowLingo has not run a matched comparison of measured and learned features on its Atlas recordings. Explore the self-supervised audio guide and model comparison for examples and their limits.
A projected map suggests questions
A projection can help select recordings to listen to and compare. Its geometry depends on the representation and projection settings. The UMAP documentation describes how neighborhood size, minimum distance and the distance metric affect the result; density preservation has a separate densMAP extension.
A bridge between points, a subcluster or a shifted centroid is a hypothesis to investigate. Establishing caller identity, dialect or behavioral context needs independent observations and a suitable evaluation. CrowLingo's Atlas is an editorial grouping of recordings; its layout is not the output of a fitted embedding model. See latent space and projections.
Compare evidence on the same question
- State the task and the population or recording conditions the result should generalize to.
- Document the measurements or exact model version, preprocessing and training data.
- Use a shared evaluation set and a split that prevents recording leakage and matches the intended generalization.
- Report the chosen metric, uncertainty and failure cases alongside listening checks and biological observations.
Measured features and learned representations can be combined. The useful choice depends on the question, available observations and demonstrated performance under the intended conditions.