This repository contains the development framework for the BioDCASE-Tiny 2025 competition (Task 3), focusing on TinyML implementation for bird species recognition on the ESP32-S3-Korvo-2 development board.

For complete competition details, visit the official BioDCASE 2025 Task 3 website.

Background

BioDCASE-Tiny is a competition for developing efficient machine learning models for bird audio recognition that can run on resource-constrained embedded devices. The project uses the ESP32-S3-Korvo-2 development board, which offers audio processing capabilities in a small form factor suitable for field deployment.

Table of Contents

• Setup and Installation
• Usage
• Dataset
• Development
  • Data Processing Pipeline
  • Model Training
  • ESP32-S3 Deployment
• ESP32-S3-Korvo-2 Development Board
• Code Structure
• Development Tips
• Evaluation Metrics
• License
• Citation
• Funding
• Partners

Setup and Installation

Prerequisites

1. Python 3.11+ with pip and venv
2. Docker for ESP-IDF environment
3. USB cable and ESP32-S3-Korvo-2 development board

Installation Steps

1. Clone the repository: bash git clone https://github.com/birdnet-team/BioDCASE-Tiny-2025.git cd BioDCASE-Tiny-2025

2. Create a virtual environment (recommended)

bash python3 -m venv .venv source .venv/bin/activate

1. Install Python dependencies: bash pip install -e .

2. Set your serial device port in the pipeline_config.yaml

yaml embedded_code_generation: serial_device: <YOUR_DEVICE>

Running on Windows

As the required tflite-micro package is not easily available for Windows we recommend using WSL to run this project.

To make your device accessible for WSL you can use this guide: https://learn.microsoft.com/en-us/windows/wsl/connect-usb

To determine your serial device port you can use the following command:

bash dmesg | grep tty

You might also need to grant some rights to run the deployment:

bash sudo adduser $USER dialout sudo chmod a+rw $SERIAL_PORT

Usage

• Modify model.py with your architecture (make sure to compile with optimizer and loss)
• Modify the training loop in the same file, if you need to
• Modify pipeline_config.yaml parameters of feature extraction
• run biodcase.py

Important: Writing custom features rather than using what is implemented here, requires implementing a numerically equivalent version on the embedded target too, as that is a necessary condition for the evaluated model to behave identically on the host and on the embedded target. This is a non-trivial undertaking, so we generally advice to stick to what is already implemented in this repository.

Dataset

The BioDCASE-Tiny 2025 competition uses a specialized dataset of Yellowhammer bird vocalizations. Key features include:

• 2+ hours of audio recordings
• Songs from multiple individuals recorded at various distances (6.5m to 200m)
• Recordings in different environments (forest and grassland)
• Includes negative samples (other bird species and background noise)
• Dataset is split into training, validation, and evaluation sets

The training set contains recordings from 8 individuals, while the validation set contains recordings from 2 individuals. An additional 2 individuals are reserved for the final evaluation.

Dataset Structure

The dataset is organized as follows:

Development_Set/ ├── Training_Set/ │ ├── Yellowhammer/ │ │ └── *.wav (Yellowhammer vocalizations) │ └── Negatives/ │ └── *.wav (Other species and background noise) └── Validation_Set/ ├── Yellowhammer/ │ └── *.wav (Yellowhammer vocalizations) └── Negatives/ └── *.wav (Other species and background noise)

• Yellowhammer/ - Contains target species vocalizations with filenames following format YH_SongID_Location_Distance.wav

  • SongID: 3-digit identifier for each song
  • Location: "forest", "grassland", or "speaker" for original recordings
  • Distance: A (original) through H (farthest at ~200m)

• Negatives/ - Contains negative samples with filenames following format Type_ID.wav

  • Type: "Background" for noise or "bird" for non-target species vocalizations
  • ID: File identifier

Download

Download the dataset from: BioDCASE-Tiny 2025 Dataset

After downloading paste the folders into /data/01_raw/clips

Development

Quickstart

To run the complete pipeline execute: bash python biodcase.py

This will execute the data preprocessing, extract the features, train the model and deploy it to your board.

Once deployed, benchmarking code on the ESP32-S3 will display info, via serial monitor, about the runtime performance of the preprocessing steps and actual model.

Step-by-Step Deployment Instructions

The steps of the pipeline can be executed individually

1. Data Preprocessing bash python data_preprocessing.py

2. Feature Extraction bash python feature_extraction.py

3. Model Training bash python model_training.py

4. Deployment bash python embedded_code_generation.py

Data Processing Pipeline

The data processing pipeline follows these steps: 1. Raw audio files are read and preprocessed 2. Features are extracted according to configuration in pipeline_config.yaml 3. The dataset is split into training/validation/testing sets 4. Features are used for model training

Model Training

The model training process is managed in model_training.py. You can customize: - Model architecture in model.py and, optionally, the training loop - Training hyperparameters in pipeline_config.yaml - Feature extraction parameters to optimize model input

ESP32-S3 Deployment

To deploy your model to the ESP32-S3-Korvo-2 board, you'll use the built-in deployment tools that handle model conversion, code generation, and flashing. The deployment process:

1. Converts your trained Keras model to TensorFlow Lite format optimized for the ESP32-S3
2. Packages your feature extraction configuration for embedded use
3. Generates C++ code that integrates with the ESP-IDF framework
4. Compiles the firmware using Docker-based ESP-IDF toolchain
5. Flashes the compiled firmware to your connected ESP32-S3-Korvo-2 board

ESP32-S3-Korvo-2 Development Board

The ESP32-S3-Korvo-2 board features: - ESP32-S3 dual-core processor - Built-in microphone array - Audio codec for high-quality audio processing - Wi-Fi and Bluetooth connectivity - USB-C connection for programming and debugging

Code Structure

Key Entry Points

• biodcase.py - Main execution pipeline
• model.py - Define your model architecture
• feature_extraction.py - Audio feature extraction implementations
• embedded_code_generation.py - ESP32 code generation utilities
• biodcase_tiny/embedded/esp_target.py - ESP target definition and configuration
• biodcase_tiny/embedded/firmware/main - Firmware source code

Benchmarking

The codebase includes performance benchmarking tools that measure: - Feature extraction time - Model inference time - Memory usage on the target device

Development Tips

1. Feature Extraction Parameters: Carefully tune the feature extraction parameters in pipeline_config.yaml for your specific audio dataset.

2. Model Size: Keep your model compact. The ESP32-S3 has limited memory, so optimize your architecture accordingly.

3. Profiling: Use the profiling tools to identify bottlenecks in your implementation.

4. Memory Management: Be mindful of memory allocation on the ESP32. Monitor the allocations reported by the firmware.

5. Docker Environment: The toolchain uses Docker to provide a consistent ESP-IDF environment, making it easier to build on any host system.

Evaluation Metrics

The BioDCASE-Tiny competition evaluates models based on multiple criteria:

Classification Performance

• Average precision: the average value of precision across all recall levels from 0 to 1.

Resource Efficiency

• Model Size: Tflite model file size (KB)
• Inference Time: Average time required for single audio classification, including feature extraction (ms)
• Peak Memory Usage: Maximum RAM usage during inference (KB)

Ranking

Participants will be ranked separately for each one of the evaluation criteria.

License

This project is licensed under the Apache License 2.0 - see the license headers in individual files for details.

Citation

If you use the BioDCASE-Tiny framework or dataset in your research, please cite the following:

Framework Citation

bibtex @misc{biodcase_tiny_2025, author = {Carmantini, Giovanni and Förstner, Friedrich and Isik, Can and Kahl, Stefan}, title = {BioDCASE-Tiny 2025: A Framework for Bird Species Recognition on Resource-Constrained Hardware}, year = {2025}, institution = {Cornell University and Chemnitz University of Technology}, type = {Software}, publisher = {GitHub}, journal = {GitHub Repository}, howpublished = {\url{https://github.com/birdnet-team/BioDCASE-Tiny-2025}} }

Dataset Citation

bibtex @dataset{yellowhammer_dataset_2025, author = {Morandi, Ilaria and Linhart, Pavel and Kwak, Minkyung and Petrusková, Tereza}, title = {BioDCASE 2025 Task 3: Bioacoustics for Tiny Hardware Development Set}, year = {2025}, publisher = {Zenodo}, doi = {10.5281/zenodo.15228365}, url = {https://doi.org/10.5281/zenodo.15228365} }

Funding

This project is supported by Jake Holshuh (Cornell class of ´69) and The Arthur Vining Davis Foundations. Our work in the K. Lisa Yang Center for Conservation Bioacoustics is made possible by the generosity of K. Lisa Yang to advance innovative conservation technologies to inspire and inform the conservation of wildlife and habitats.

The development of BirdNET is supported by the German Federal Ministry of Education and Research through the project “BirdNET+” (FKZ 01|S22072). The German Federal Ministry for the Environment, Nature Conservation and Nuclear Safety contributes through the “DeepBirdDetect” project (FKZ 67KI31040E). In addition, the Deutsche Bundesstiftung Umwelt supports BirdNET through the project “RangerSound” (project 39263/01).

Partners

BirdNET is a joint effort of partners from academia and industry. Without these partnerships, this project would not have been possible. Thank you!