TinyML‐Based Burrowing Owl Acoustic Detection

Abstract

A keystone species, burrowing owls convey crucial ecological information through their vocalizations. This project creates a whole TinyML pipeline that listens for, recognizes, and categorizes six different burrowing-owl sounds in real time, all of which are controlled by an STM32H7 microcontroller. To train lightweight CNNs (Custom Tiny CNN, MobileNetV2, ProxylessNAS) on the BUOWSET dataset, we transform raw audio into 64-band Mel spectrograms, quantize the model to int8, and embed it as a C header for on-device inference.

Project Overview

Our goal is to build a TinyML-based system that runs entirely on an STM32H7 board to listen for, detect, and classify six different burrowing-owl vocalizations in real time. By converting audio to 64-band Mel spectrograms, training lightweight CNNs (MobileNetV2 and ProxylessNAS) on the BUOWSET dataset, then quantizing to int8 and embedding the model as a C header, we achieve:

≥ 90 % classification accuracy on the BUOWSET test set
Flash/RAM footprint ≤ 0.5 MB after quantization
On-chip Mel-filterbank DSP and inference loop
Optimized inference code generated via X-Cube-AI for STM32

Models Used

We trained three compact CNN architectures optimized for on-device inference on STM32H7. Final validation accuracy and F1 score (epoch 20) are shown below:

Custom Tiny CNN — Accuracy: 94.8 %, F1 Score: 84.0 % , Size = 33 KB (Pre-quantization)
MobileNetV2 — Accuracy: 97.7 %, F1 Score: 91.8 % , Size = 9.2 MB (Pre-quantization)
ProxylessNAS — Accuracy: 97.3 %, F1 Score: 91.9 % , Size = 11.5 MB (Pre-quantization)

Training Curves

Custom Tiny CNN: Train & Val Loss

Custom Tiny CNN Training vs Validation Loss

Custom Tiny CNN: Precision, Recall, F1

Custom Tiny CNN Precision, Recall, and F1 over Epochs

MobileNetV2: Train & Val Loss

MobileNetV2: Validation Metrics

MobileNetV2 Validation Accuracy, Precision, Recall, F1

ProxylessNAS: Train & Val Loss

ProxylessNAS Training vs Validation Loss

ProxylessNAS: Validation Metrics

ProxylessNAS Validation Accuracy, Precision, Recall, F1

Quick Links & Resources

Hardware & Software Repositories

Hardware STM32 DSP Code: STM32-Acoustic-Species-Hardware
Software ML & Quantization: TinyML-Owl-Acoustic-Species

TinyML Video Presentation

Documentation & Reports

Media & Slides

Model Demo Video (YouTube): https://youtu.be/L2aSu9MTcUA
Class Slides (PDF): Slides Link

Wiki & README

GitHub README (Documentation):
ZhenmanShen/CSE145-237D-ML-README.md

Team Members

Software

Zach Lawrence
zclawrence@ucsd.edu
Abhay Lal
ablal@ucsd.edu
Max Shen
zhs009@ucsd.edu

Hardware

Kruti Dharanipathi
kdharanipathi@ucsd.edu
Reese Whitlock
rwhitlock@ucsd.edu
Ben Scott
bmscott@ucsd.edu

Acknowledgements

We gratefully acknowledge the support and data provided by the Engineers For Exploration (E4E) Acoustic Species Identification Lab at UC San Diego. E4E is a research group focused on protecting the environment, uncovering mysteries related to cultural heritage, and providing experiential learning experiences for undergraduate and graduate students.

This work was conducted in collaboration with the San Diego Zoo Wildlife Alliance, whose expertise in burrowing-owl ecology and field data collection was invaluable.

We also thank Professor Ryan Kastner for his guidance and support throughout this project.

Contact & Project Lead at E4E:
Email: lvonschoenfeldt@ucsd.edu