Skip to content
CrowLingo
Back to the pipeline

Stage 05

Cluster & label — test the groups before naming them.

Stages 5–7 connect representations, descriptive labels and observed context. A cluster is a result of analytical choices; a compelling map does not establish a crow call's meaning.

Stages 5–7 — Project, label and contextualize

Clustering methods can suggest groups in measured acoustic representations. Group counts and labels require validation against recordings and observations. CrowLingo’s nine atlas groups are editorial categories assigned from source metadata, not HDBSCAN results.

Audio narration is unavailable while its source claims are corrected.

Choose a space for the question

A two-dimensional projection can make recordings easier to inspect, but it can distort distances and apparent group boundaries. The UMAP authors also describe dimensionality reduction for clustering, including with HDBSCAN. Clustering need not always use either the full embedding space or a two-dimensional display.

Justify the representation, dimension and metric for your task. Compare plausible choices, check stability and inspect the underlying recordings. A projection that looks convincing is not validation. See Latent space & UMAP for the distinction between a representation and a visualization.

Treat parameters as choices to evaluate

HDBSCAN is one possible grouping method, not an established default for this crow corpus. This guide does not prescribe a minimum cluster size or a target percentage of grouped recordings. Those choices require a stated task, data inspection and validation.

Preserve the clip-to-vector mapping from the embedding stage. Record the algorithm and software version, metric, parameters, random seeds where applicable, assignments and unassigned examples. Report how conclusions change under reasonable alternative settings.

Label what you can actually observe

Listen to representative and atypical recordings from each proposed group, inspect their spectrograms, and document how examples were selected. Use provisional acoustic descriptions such as “repeated harsh caws.” A descriptive label alone does not establish territory defense or another behavioral context.

Check proposed labels against independent annotations or observations appropriate to the claim. At stage 7, synchronized behavioral records can support testing a context association. Keep that evidence separate from the acoustic grouping and retain uncertainty.

Keep the difficult examples visible

Unassigned recordings do not automatically indicate dialect, graded variation or recording noise. Inspect them rather than assigning an explanation from their position on a plot. Report exclusions and unresolved cases; there is no validated 80/20 coverage target for this corpus.

Adding recordings or changing a model can change assignments and cluster identifiers. Preserve versioned runs and an explicit mapping when comparing them. A centroid or numeric ID is not a stable biological identity.