BirdNET-Pi Vocalization Classifier

Classify bird vocalizations as song, call, or alarm

Works with BirdNET-Pi to add vocalization context to your bird detections.

What does it do?

Detection	Without	With Vocalization
Eurasian Blackbird	"Merel detected"	"Merel - Zang (93%)"
European Robin	"Roodborst detected"	"Roodborst - Alarm (87%)"

Why is this useful?

• Song: Bird is marking territory or attracting mate
• Call: Contact calls, flock communication
• Alarm: Predator nearby! (cat, sparrowhawk, etc.)

Quick Start

Download Pre-trained Models

197 Ultimate models trained on Google Colab A100 are available for download:

📥 Download from Google Drive (~6.9 GB total)

Individual models are ~35 MB each. Download only the species you need, or get them all.

Use the Classifier

from src.classifiers.cnn_inference import VocalizationClassifier

classifier = VocalizationClassifier(models_dir="./models")
result = classifier.classify("Koolmees", "/path/to/audio.mp3")

if result:
    print(f"{result['type']} ({result['confidence']:.0%})")
    # Output: song (91%)

Train Your Own Models

# Clone the repository
git clone https://github.com/RonnyCHL/emsn-vocalization.git
cd emsn-vocalization

# Install dependencies
pip install torch librosa numpy scikit-learn tqdm requests

# Train a model (downloads data from Xeno-canto automatically)
python train_existing.py --species "Koolmees"

Train on Google Colab (Free GPU)

Open notebooks/EMSN_Vocalization_Colab_Training.ipynb in Google Colab for free GPU training.

How it Works

┌─────────────────┐     ┌──────────────────┐     ┌─────────────────┐
│  BirdNET-Pi     │     │  Vocalization    │     │   Result        │
│  "Merel"        │ ──▶ │  Classifier      │ ──▶ │  "Merel - Zang" │
│                 │     │  (CNN model)     │     │  (93%)          │
└─────────────────┘     └──────────────────┘     └─────────────────┘

1. BirdNET-Pi identifies the bird species from audio
2. This classifier analyzes the same audio with a species-specific CNN
3. Output includes vocalization type (song/call/alarm) with confidence

Project Structure

emsn-vocalization/
├── src/
│   ├── classifiers/      # CNN model & inference
│   ├── collectors/       # Xeno-canto data collection
│   └── processors/       # Audio → spectrogram processing
├── notebooks/            # Colab training notebooks
├── train_existing.py     # Main training script
├── full_pipeline.py      # Complete pipeline (download → train)
└── docker-compose.yml    # Docker training environment

Model Details

Ultimate Models (recommended)

• Architecture: 4-layer CNN with batch normalization (32→64→128→256 filters)
• Classifier: 512→256→num_classes with dropout
• Training: Google Colab A100, 50 epochs, data augmentation
• Size: ~35 MB per species model
• Accuracy: Improved over standard models

Standard Models

• Architecture: 3-layer CNN (32→64→128 filters)
• Classifier: 256→num_classes
• Size: ~2 MB per species model

Common specs

• Input: Mel spectrograms (128x128, 3 seconds audio)
• Output: song / call / alarm + confidence
• Sample rate: 48 kHz, freq range: 500-8000 Hz

Available Models

Currently 197 trained models for Dutch bird species, including:

• Koolmees (Great Tit)
• Merel (Eurasian Blackbird)
• Roodborst (European Robin)
• Huismus (House Sparrow)
• Vink (Common Chaffinch)
• ... and 192 more

See the Google Drive folder for the complete list.

Integration Options

Standalone (recommended for testing)

Run as separate service, reads BirdNET-Pi's birds.db.

With BirdNET-Pi

Can be integrated to show vocalization in the web interface.

See COMMUNITY_PITCH.md for integration discussion.

Requirements

• Python 3.10+
• PyTorch 2.0+
• librosa
• numpy
• Raspberry Pi 4/5 (for inference) or any Linux system

Training Data

Audio data is automatically downloaded from Xeno-canto:

• Quality A/B recordings preferred
• Balanced sampling across vocalization types
• Respects Xeno-canto API rate limits

Contributing

• Test the classifier: Try it with your BirdNET-Pi setup
• Train more species: Use Colab notebook to train new models
• Report issues: Open a GitHub issue
• Integration ideas: See community pitch document

Related Projects

• BirdNET-Pi - Bird species identification
• Xeno-canto - Bird sound database

License

MIT License - free to use, modify, and distribute.

Author

Ronny Hullegie - EMSN Project (Ecologisch Monitoring Systeem Nijverdal)

Citation

@software{hullegie2025vocalization,
  author = {Hullegie, Ronny},
  title = {BirdNET-Pi Vocalization Classifier},
  year = {2025},
  url = {https://github.com/RonnyCHL/emsn-vocalization}
}