Desenvolvimento de um sistema embarcado para monitoramento bioacústico da fauna com TinyML

Abstract

This project presents the development of a compact device capable of capturing environmental sounds and identifying bird vocalizations using embedded machine learning system. The system was built with an ESP32 microcontroller and an I2S digital microphone, allowing all audio processing to be performed directly on the hardware, without relying on internet access or external services. The implementation of the system involved two main stages. The first consisted of gathering audio recordings and organizing a dataset containing bird vocalizations and environmental sounds. These samples were processed and used to train a classification model on the Edge Impulse platform. Once the final model was obtained, it was converted into an optimized version compatible with microcontrollers TinyML. In the second stage, the trained model was deployed to the ESP32, enabling real-time analysis of the recorded audio. Whenever a segment of sound is captured, the device performs preprocessing, extracts the required features, and applies the classifier. The results are then stored on a microSD card, allowing later review and evaluation. The proposed system demonstrates that low-power microcontrollers can be used for acoustic recognition tasks in the field, with continuous operation and minimal energy consumption. This approach can support environmental research, wildlife monitoring, and conservation studies, as the device can remain in remote areas without the need for constant supervision.