Relational acoustic geometry of over 20,000 common nightingale strophes: a framework for continuous vocal analysis

Lucio Arese 

Large collections of complex animal vocalizations are often organized through predefined song categories, which provide powerful descriptive units but do not directly characterize continuous and representation-dependent relationships within the acoustic material. This study presents a field-grounded framework for examining such relationships without assigning vocalizations to categories in advance. During a 68-day field campaign in the Cuneo Fluvial Park, Italy, 227 georeferenced recordings comprising approximately 42.5 h of audio were collected across ten operational observation areas. Of these, 161 recordings were retained, yielding 20,075 indexed nightingale strophes. Each strophe was represented in parallel as high-dimensional MFCC-40 and microtonal Chroma-80 time–feature matrices, referred to here as acoustic terrains, and projected into separate globally fitted PCA spaces. Three-dimensional UMAP embeddings supported interactive navigation and acoustic inspection, whereas quantitative analyses were conducted in the retained PCA coordinates using observation-area centroids, within-area dispersion, and exact nearest-neighbor relationships under Euclidean and cosine metrics.

Strophes from the ten observation areas occupied broadly shared but non-uniform acoustic domains. Their organization was expressed through differences in centroid position and dispersion and through directional enrichment within local neighborhoods, rather than through isolated source-specific clusters. MFCC generally retained stronger same-area concentration than chroma, while Euclidean and cosine distances produced substantially different neighborhood compositions. Relational organization was therefore conditional both on the acoustic properties emphasized by the representation and on the geometrical criterion used to define proximity. Observation-area-associated structure remained measurable after candidates from the focal recording were excluded, and the broader ordering of directional enrichment relationships was substantially preserved across neighborhood sizes and under equal-contribution subsampling, although their magnitude remained sensitive to corpus composition and analytical scale. Denoising sensitivity analyses additionally showed high directional agreement between feature profiles before and after the final denoising pass.

Because observation areas represent operational field units rather than confirmed individual identities, these relationships should not be interpreted as definitive individual repertoires, conventional strophe types, or functional vocal classes. The study instead establishes a documented corpus and a reproducible framework in which relationships among complete vocal events can be examined before categorical assignment, providing a basis for subsequent categorical, sequential, temporal, and longitudinal investigation of nightingale song and for empirical evaluation in other complex animal vocal systems.