Acoustic Environmental Sensing on the Edge

This paper presents a next-gen tinyML Smart Weather Station for Acoustic En-vironmental Sensing (AES). We present a low cost, low power, reliable acousticenvironmental sensor that provides the functions of a next generation weatherstation but is extensible to capturing environmental noise as well. This version ofthe sensor captures temperature, humidity and atmospheric pressure readings(TPHdata) as well as the intensity of rain derived from a tiny ML model that calculatesthe intensity of the rain from recorded environmental sound. We present the designof a 3D printed robust sensor using a custom design based on the Raspberry Pi 3B+fused with an INMP441 MEMS microphone that integrates over I2S, with a pricetag of $125. We also present a tiny ML TFLite single headed model customizedfor our context that outputs rain intensity but is extensible to other environmentalnoises. We show that our models based on the YAMNet feature extraction modelsprovide superior performance and operate efficiently on the low resources of thesensor. The model operates at an accuracy of 93% on the test set making it versatilefor weather and noise classification tasks effectively while maintaining a small sizeof 516Kb.