BioDCASE 2026 Task 3: Bioacoustics for Tiny Hardware Development Set

Overview

This dataset contains annotated avian acoustic recordings tailored for the BioDCASE 2026 TinyML Task 3 challenge. It is designed to facilitate the development and evaluation of lightweight machine learning models capable of identifying bird species from audio recordings in resource-constrained environments.

The dataset includes recordings of several common European bird species as well as a dedicated background noise class representing urban/town environments. Original recordings were made at various sites in Germany using eciPi recording units by OekoFor, and all samples have been carefully annotated by Ralph Martin to ensure high-quality call centers.

Technical Specifications

• Format: .wav
• Sample Rate: 24,000 Hz
• Channels: 1 (Mono)
• Bit Depth: 16-bit PCM
• Duration: 3.0 seconds per snippet
• Total Runtime (Development Set): ~2.29 hours

Categories

The dataset consists of 10 bird species plus 1 background class (11 classes in total).

(Note: The exact list of 10 species depends on the specific source selection but generally includes typical garden/forest birds such as Common Chaffinch, Eurasian Blackbird, Great Tit, etc., as well as a 'Background' class).

Dataset Structure

The dataset provided here is the Development Set (for training and validation). A separate, balanced Test Set is held out and kept hidden for the final challenge evaluation. The distribution per class in the development set is:

• Development Set (BioDCASE2026_TinyML_Development_Dataset.zip)
  • Train: 200 samples per class
  •  Validation: 50 samples per class

 (Note: A hidden/hold-out test set containing 50 samples per class is maintained separately for scoring challenge submissions).

Each split contains subfolders corresponding to the species/class name.

Folder Structure Example

Development_Set/
├── Train/
│   ├── Background/
│   ├── Common_Chaffinch/
│   ├── ...
├── Validation/
│   ├── Background/
│   ├── Common_Chaffinch/
│   ├── ...

File Naming Convention

Files are systematically named to easily identify their split and ground-truth species:

BioDCASE2026_TinyML_[SPLIT]_[ID]_[Species].wav

• [SPLIT]: TRAIN, VAL, or TEST
• [ID]: A 4-digit sequential identifier (e.g., 0001)
• [Species]: The formatted species name (e.g., Common_Chaffinch)

Example: BioDCASE2026_TinyML_TRAIN_0042_Common_Chaffinch.wav

License

This dataset is published under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license.