Bird Call Classifier for Solar Audio Logging

ISEF Category: Embedded Systems

Subcategory: Signal Processing  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Bird sounds can hide in plain noise. A tiny chip can still tell species apart if you give it the right features and a model that fits its memory. That means your project can sit on a tree, run on solar power, and make decisions without a laptop. You are not just detecting sound, you are building an edge AI system.

What Is It?

This project asks a simple question with a hard answer, can a tiny microcontroller identify bird species from short audio clips? The chip cannot store a huge model or do heavy math, so you have to shrink the problem. One common trick is to turn sound into a log-Mel spectrogram, which is a picture of how sound energy changes across pitch and time. Then a small neural network, like MobileNetV3-tiny, learns patterns in that picture.

Think of it like teaching a fast, tired student to recognize birds from quick sketches instead of full-color photos. The sketches keep the important shape, but leave out extra detail. Your job is to test how much detail the model really needs, and how far you can push accuracy while staying inside the ESP32-S3’s memory, compute, and power limits. The solar logger part adds a second layer, because the device also has to survive in the field for long periods.

Why This Is a Good Topic

This is a strong science fair topic because you can test many parts of the system one by one. You can compare feature types, model sizes, compression methods, and power budgets with real metrics like accuracy, latency, memory use, and energy per classification. The project connects to wildlife monitoring, which matters for conservation and habitat studies. You can learn signal processing, machine learning, embedded systems, and experimental design in one project.

Research Questions

• How does the size of the MobileNetV3-tiny model affect bird-call classification accuracy on an ESP32-S3?
• What is the effect of different log-Mel spectrogram settings on species-level accuracy?
• Does quantizing the model to lower precision change accuracy more than it reduces memory use?
• To what extent does background noise reduce classification confidence for common BirdCLEF species?
• Which audio clip length gives the best tradeoff between latency, energy use, and accuracy?
• How does solar charging variability affect the total number of calls the logger can process per day?

Basic Materials

• ESP32-S3 development board with microphone input support.
• MEMS microphone module or I2S microphone.
• Solar panel sized for low-power electronics testing.
• Rechargeable lithium battery pack with protection circuit.
• Breadboard and jumper wires.
• USB cable and computer.
• Digital multimeter.
• Notebook or spreadsheet for test logs.
• Headphones or small speaker for audio checks.
• Access to labeled bird audio clips from BirdCLEF or a similar open dataset.

Advanced Materials

• ESP32-S3 board with external PSRAM.
• MEMS microphone with known frequency response.
• Solar charger module and power management board.
• Programmable electronic load or power analyzer.
• Oscilloscope for checking audio and power rails.
• Calibrated reference speaker or audio source.
• MicroSD module for on-device logging.
• Temperature and light sensor for field condition logging.
• Access to a test enclosure for outdoor deployment.
• Labeled BirdCLEF training and validation data.

Software & Tools

• Python: Prepares audio features, trains baseline models, and analyzes accuracy and power results.
• Jupyter Notebook: Helps you compare feature settings, plots, and confusion matrices in one place.
• TensorFlow Lite Micro: Runs a compressed classifier on the ESP32-S3.
• Audacity: Checks audio clips, trims samples, and inspects noise before feature extraction.
• ImageJ: Can inspect spectrogram exports as images if you need a quick visual comparison.

Experiment Steps

1. Define the exact bird-call classes and the field setting you want to support.
2. Choose one target metric first, such as accuracy, latency, memory use, or energy per call.
3. Build a baseline offline model so you know what performance looks like before compression.
4. Decide which feature pipeline you will compare, then keep the rest of the pipeline fixed.
5. Plan a deployment test that measures both recognition quality and power budget on the ESP32-S3.
6. Set controls for noise, distance, and clip length so you can separate signal quality from hardware limits.

Common Pitfalls

• Training on clips that contain the same recording location in both train and test splits, which inflates accuracy.
• Ignoring class imbalance, which makes the model favor common bird species and miss rare ones.
• Changing the audio preprocessing between laptop training and ESP32 deployment, which causes a drop in real-world performance.
• Measuring power only during active inference, which hides the cost of wake-up, capture, and storage.
• Testing only clean clips, which makes the logger look better than it will in wind, rain, or insect noise.

What Makes This Competitive

A competitive project will go beyond asking whether the model works. You should compare feature pipelines, model sizes, and power budgets with the same test split and the same hardware constraints. Strong entries often include ablation tests, confusion analysis by species, and a clean tradeoff curve for accuracy versus energy use. If you also test field noise or solar variability, your project starts to look like a real deployment study, not just a demo.

Project Variations

• Compare log-Mel spectrograms with MFCC features on the same bird-call dataset and measure which one compresses better for the ESP32-S3.
• Test how well the logger detects one local bird family, such as warblers or sparrows, instead of many species at once.
• Swap the solar-powered setup for a battery-only setup and analyze how duty cycle changes classification throughput.

Learn More

• BirdCLEF, Kaggle competition archive: Search Kaggle for BirdCLEF datasets and baselines to find labeled bird audio for model training and evaluation.
• A Course in Machine Learning: Search for a free university-style machine learning textbook to review feature extraction, classification, and model validation.
• TensorFlow Lite Micro documentation: Read the official guides for deploying small neural networks on microcontrollers.
• ESP32-S3 technical reference manual, Espressif: Search the official Espressif documentation for memory limits, audio interfaces, and low-power features.
• NOAA National Centers for Environmental Information: Use weather and solar data to connect your field deployment plan to real environmental conditions.
• PubMed: Search review articles on acoustic monitoring and passive wildlife survey methods.