BirdCLEF+ 2026 Workspace

This repository contains local experiments, preprocessing scripts, and notebook workflows for the BirdCLEF+ 2026 Kaggle competition.

The current workflow focuses on:

• generating mel-spectrogram PNG datasets from competition audio,
• packaging ranked species batches into zip archives for Kaggle dataset upload,
• training image models (fastai/ResNet baseline) on spectrograms.

Competition links

• Competition home: BirdCLEF+ 2026
• Data page: Competition data
• Overview: Problem statement and context
• Rules: Submission and usage rules
• Leaderboard: Public standings
• Code tab: Community notebooks
• Discussion tab: Q&A and competition updates

Repo contents

• scripts/generate_spectrogram_batches.py
  Generates spectrogram images from train_audio, batches species by frequency rank, zips each batch, and writes manifest files.
• artifacts/spectrogram_batches/
  Output folder for generated batch zips and manifests.
• birdclef_plus_2026_sound_classification_attempt_1.ipynb
  Main Kaggle-oriented training notebook (spectrogram image classifier experiments).
• data/birdclef-2026/
  Local extracted competition files (train.csv, train_audio/, train_soundscapes/, etc.).

Environment

This project uses uv with the repo virtual environment:

source .venv/bin/activate
uv sync

Run scripts with uv run, for example:

uv run python scripts/generate_spectrogram_batches.py

Notes on labels and data splits

• taxonomy.csv and sample_submission.csv contain 234 classes.
• train.csv currently has labeled training clips for 206 classes.
• Additional classes can be represented in train_soundscapes_labels.csv via labeled time segments in train_soundscapes/.