Edge-Based Long-Term Automated Monitoring of Budgerigar Vocal Activity Using YAMNet

Continuous monitoring of animal behavior may provide quantitative information on responses to environmental changes; however, its application in seismobiological research requires reliable characterization of baseline behavioral variability. This study developed and evaluated an edge-based bioacoustic monitoring system for the continuous analysis of vocal activity in budgerigars (Melopsittacus undulatus). Ten birds were monitored under stable housing conditions from June 2025 to May 2026. Audio was processed locally on a Raspberry Pi-based system that combined a pretrained YAMNet feature extractor with a task-specific binary classifier. Vocal activity was quantified using the vocal activity second (VAS), defined as a 1 s interval containing at least one detected vocalization. Time-series analysis revealed a pronounced diel organization of vocal activity, with increased activity during daylight hours and minimal activity at night, together with seasonal variability. Wavelet and periodogram analyses identified a periodic component near 24 h, with weaker components within the 2-8 h range. These findings demonstrate the feasibility of autonomous edge-based monitoring of avian vocal behavior and establish a quantitative analysis of acoustic activity. The present study represents a methodological framework for the collection and analysis of avian acoustic signals using machine learning and computational signal processing approaches. The proposed approach provides a quantitative framework for investigating changes in avian behavior and for evaluating whether potential behavioral responses in vocal organization are associated with environmental factors.