SpectroBird: Event-Level Vocalization Detection and Curation for Avian Ecology
Passive acoustic monitoring often summarizes long recordings as fixed-window presence predictions, but many ecological questions require records of individual vocal events: how many occurred, when they occurred, whether they overlapped, how long they lasted, what frequency band they occupied, and whether each event was complete enough for structural measurement. We address this event-level measurement problem with SpectroBird, a spectrogram-based method for localizing and curating avian vocalizations in noisy field recordings. SpectroBird first applies a detection-localization model to identify candidate vocalizations as time-frequency boxes. An event-consolidation step merges repeated detections from overlapping windows while preserving overlapping vocalizations as separate events. A completeness-curation model then classifies each consolidated event as complete or incomplete before downstream event-level analysis. We demonstrate SpectroBird on breeding-season Hume's Leaf Warbler buzz vocalizations recorded in the Himalaya across a 3,200-3,800 m elevational gradient. Under shared window-level scoring, the detection-localization model remained competitive with AST and YOLOv11-cls baselines while retaining event timing and frequency bounds; event-level evaluation achieved AP > 0.999 for complete buzzes. Applied to two field seasons, the resulting event records showed that duration of complete buzzes declined with elevation, complete-buzz occurrence showed no simple elevation trend, and relative day-of-year occurrence patterns were strongly correlated between years. These results show how event-level detection and curation can support ecological analyses that depend on individual vocalizations rather than fixed-window presence alone.