On-site acoustic identification of bird species based on a shallow neural network

Passive acoustic monitoring is a powerful tool for biodiversity assessment, but it often requires collecting massive amounts of data that are processed off-site, a procedure that is costly, time-intensive, and generates a significant carbon footprint. To address this, we present a method for on-site acoustic identification of bird species using a shallow neural network (NN) embedded in an AudioMoth recording unit. Focusing on the lesser kestrel ( Falco naumanni ) as a model species, the proposed algorithm leverages a shallow NN with a single hidden layer to classify 32-ms audio segments. The network processes a feature vector composed of 12 Mel Frequency Cepstral Coefficients (MFCCs) and their corresponding delta coefficients to determine the presence or absence of the target species. The complete identification algorithm was successfully integrated in the AudioMoth firmware, enabling real-time, on-site analysis without the need for additional hardware. The system demonstrated similar performance to that of BirdNet in identifying lesser kestrel’s calls in field tests, reaching a maximum F1-score of 0.89. Power consumption measurements are also reported. The proposed methodology is general and can be adapted for other species. • A neural network was implemented directly on the popular AudioMoth device. • The embedded model achieves real-time inference with only a 30% power increase. • Lesser kestrel calls were identified on AudioMoth units deployed in the field with a maximum F1-score of 0.89. • The inference performance of the proposed model is similar to that of BirdNET while fully operating on-device. • The code and datasets are publicly available and the method is generalizable to other species and platforms.