Birdcall-USA

CNN vocalization classifier for North American birds - classifies songs, calls & alarms using spectrograms.

Download Pre-trained Models

Download v1.0 Models (70 MB) - 46 species, trained December 2024

Extract to trained-models/ folder.

Overview

This project trains deep learning models to classify bird vocalizations into three categories:

• Song - Territorial/mating vocalizations
• Call - Contact/communication calls
• Alarm - Warning/distress calls

Currently supports 50 common North American species based on eBird frequency data.

Quick Start (Google Colab)

1. Open the notebook in Google Colab (Pro+ recommended for A100 GPU)
2. Run all cells to train models
3. Models are saved to data/models/

Local Installation

git clone https://github.com/RonnyCHL/birdcall-usa.git
cd birdcall-usa
pip install torch torchaudio librosa scikit-learn matplotlib seaborn requests

Usage

Train all species

python full_pipeline.py

Train only most common species (priority 1)

python full_pipeline.py --priority 1

Train a specific species

python full_pipeline.py --species "American Robin"

Dry run (see what would be trained)

python full_pipeline.py --dry-run

Species List

Priority 1 - Very Common (13 species)

Species	Scientific Name	eBird Freq
Mourning Dove	Zenaida macroura	35%
Northern Cardinal	Cardinalis cardinalis	34%
American Robin	Turdus migratorius	33%
American Crow	Corvus brachyrhynchos	32%
Blue Jay	Cyanocitta cristata	28%
Song Sparrow	Melospiza melodia	25%
Red-winged Blackbird	Agelaius phoeniceus	25%
European Starling	Sturnus vulgaris	25%
American Goldfinch	Spinus tristis	24%
Canada Goose	Branta canadensis	23%
House Finch	Haemorhous mexicanus	23%
Downy Woodpecker	Dryobates pubescens	23%
Mallard	Anas platyrhynchos	22%

Priority 2 - Common (13 species)

Red-bellied Woodpecker, House Sparrow, Turkey Vulture, Black-capped Chickadee, Tufted Titmouse, Dark-eyed Junco, White-breasted Nuthatch, Northern Flicker, Great Blue Heron, Northern Mockingbird, Carolina Wren, Red-tailed Hawk, Common Grackle

Priority 3 - Regular (24 species)

Barn Swallow, Yellow-rumped Warbler, Ring-billed Gull, Gray Catbird, Common Yellowthroat, Brown-headed Cowbird, Chipping Sparrow, Tree Swallow, Eastern Bluebird, White-throated Sparrow, Killdeer, Eastern Phoebe, Cedar Waxwing, Ruby-throated Hummingbird, Baltimore Oriole, Indigo Bunting, Eastern Towhee, Brown Thrasher, Purple Finch, Pine Warbler, Carolina Chickadee, Eastern Meadowlark, Wood Thrush, Scarlet Tanager

Model Architecture

• Input: 128x128 mel-spectrogram
• Architecture: 4-layer CNN with batch normalization
• Output: 3 classes (song, call, alarm)
• Training: ~150 samples per class from Xeno-canto

Data Source

Audio recordings are automatically downloaded from Xeno-canto, the world's largest collection of bird sounds.

Project Structure

birdcall-usa/
├── data/
│   ├── raw/              # Downloaded audio files
│   ├── models/           # Trained .pt model files
│   └── spectrograms-*/   # Generated spectrograms
├── logs/                 # Training logs and plots
├── notebooks/
│   └── train_colab.ipynb # Colab training notebook
├── src/
│   ├── classifiers/      # CNN model code
│   ├── collectors/       # Xeno-canto downloader
│   └── processors/       # Spectrogram generator
├── full_pipeline.py      # Main training script
└── us_bird_species.py    # Species list

Training Time Estimates

Scope	Species	Time (A100)
Priority 1	13	~1 hour
Priority 1+2	26	~2-3 hours
All	50	~4-6 hours

Related Projects

• emsn-vocalization - Original Dutch/European version
• BirdNET - Species identification model

License

MIT License - see LICENSE file.

Acknowledgments

• Audio data from Xeno-canto contributors
• Species frequency data from eBird