BirdCLEF 2025: Bird Song Classification

Project Overview

This project focuses on the BirdCLEF 2025 competition, a machine learning challenge for bird song classification. The goal is to develop algorithms that can identify bird species from audio recordings, with applications in biodiversity monitoring and ecological research.

Data Structure

The dataset consists of several components:

Training Data

• train_audio/: Individual bird sound recordings

  • Clean, single-species recordings
  • Labeled with species information
  • Format: .ogg files at 32 kHz
  • Filenames: [collection][file_id_in_collection].ogg

• train_soundscapes/: Full 1-minute environmental recordings

  • Contains background noise and multiple species
  • Similar format to test data
  • Filenames: [site]_[date]_[local_time].ogg

Metadata Files

• train.csv: Contains metadata for training recordings

  • primary_label: Species code
  • secondary_labels: Additional species in recording
  • latitude & longitude: Recording location
  • author: Recording provider
  • rating: Quality rating (1-5)
  • collection: Source collection (XC, iNat, or CSA)

• taxonomy.csv: Species information

  • iNaturalist taxon ID
  • Class name (Aves, Amphibia, Mammalia, Insecta)

Audio Processing Pipeline

Our preprocessing pipeline follows the BirdNET paper approach:

1. Spectrogram Generation

   • Mel-spectrograms with 64 bands
   • Frequency range: 150 Hz to 15 kHz
   • FFT window size: ~32ms at 32kHz
   • 25% overlap between frames

2. Data Augmentation

   • Frequency shifts
   • Time shifts
   • Spectrogram warping
   • Ambient noise addition

3. Signal Processing

   • 3-second chunks
   • Signal strength-based detection
   • Log scaling for magnitude

To test the project

create a uv environment

uv venv

then activate it

source .venv/bin/activate

install the necessary packages from the pyproject.toml

uv pip install -r pyproject.toml

and finally run the scripts

python preprocessing.py
python augmentation.py
python training.py

References

• BirdNET Paper
• BirdCLEF 2025 Competition

License

Licensed under a MIT License.