Efficient Bird Sound Detection on the Bela Embedded System

Monitoring wildlife is an important aspect of conservation initiatives. Deep learning detectors can help with this, although it is not yet clear whether they can run efficiently on an embedded system in the wild. This paper proposes an automatic detection algorithm for the Bela embedded Linux device for wildlife monitoring. The algorithm achieves good quality recognition, efficiently running on continuously streamed data on a commercially available platform. The program is capable of computing on-board detection using convolutional neural networks (CNNs) with an AUC score of 82.5% on the testing set of an international data challenge. This paper details how the model is exported to work on the Bela Mini in C++, with the spectrogram generation and the implementation of the feed-forward network, and evaluates its performance on the Bird Audio Detection challenge 2018 DCASE data.