Enabling Multi-Species Bird Classification on Low-Power Bioacoustic Loggers

Passive acoustic monitoring is widely used in ecological research to collect and store large volumes of environmental audio for biodiversity assessment and species monitoring. A central challenge is moving from passive recording alone to in-situ inference, especially on low-power bioacoustic loggers with strict limits on memory, computation, and energy. The main goal of this thesis is to design an efficient bird sound classification pipeline that remains both accurate and deployable on microcontroller-class hardware. The proposed system, named WrenNet, combines a compact neural architecture with a semi-learnable spectral front-end that preserves an interpretable filterbank structure while allowing limited adaptation of frequency allocation. Training uses augmentation and knowledge distillation to improve robustness to noise, class imbalance, and acoustically similar species. The system is evaluated on a large dataset of bird recordings and environmental sounds. Beyond offline metrics, the model is deployed and tested on AudioMoth hardware, where latency, memory footprint, and energy consumption are measured. Results show that the proposed approach achieves effective multi-species classification while remaining compatible with the strict resource constraints of low-power bioacoustic loggers, improving scalable and energy-efficient bioacoustic monitoring.