BirdCLEF 2026 - Bird Sound Recognition System

A deep learning-based bird species recognition system using bird audio recordings and mel spectrogram classification.

---

Project Overview

This project identifies bird species from their sound recordings using:

• Audio preprocessing
• Mel spectrogram generation
• Deep learning classification
• Transfer learning with EfficientNet

The system converts bird audio into spectrogram images and classifies them into bird species categories.

---

Features

• Bird audio loading with Librosa
• Mel spectrogram generation
• EfficientNet-based CNN classifier
• Audio augmentation
• Training + validation pipeline
• Model inference pipeline
• GPU-ready training support

---

Dataset Structure

data/
│
├── train_audio/
├── train_soundscapes/
├── test_soundscapes/
│
├── train.csv
├── taxonomy.csv
├── sample_submission.csv
└── train_soundscapes_labels.csv

---

Tech Stack

• Python
• PyTorch
• Librosa
• TIMM
• Audiomentations
• Scikit-learn

---

Model Architecture

• EfficientNet-B0
• Mel Spectrogram Input
• Transfer Learning
• Dropout Regularization

---

Training Pipeline

Audio
↓
Mel Spectrogram
↓
Normalization
↓
EfficientNet
↓
Classification

---

Installation

pip install -r requirements.txt

---

Run Training

python src/train.py

---

Run Inference

python src/inference.py

---

Current Progress

• Working preprocessing pipeline
• Spectrogram visualization
• EfficientNet training
• Validation support
• Audio augmentation
• GPU migration planned

---

Future Improvements

• Full dataset training
• Multi-label classification
• AST / BirdNET integration
• Soundscape inference
• Competition submission pipeline

---

Author

Pari Jain