Bird Vocalization Event Detector (CRNN)

Author: Mark Slipenkyi

Date: 2025

Task: Multi-label audio classification and time-frequency event detection.

📌 Project Overview

This project implements a Convolutional Recurrent Neural Network (CRNN) to identify and localize bird species in audio recordings. The model processes log-mel spectrograms to provide:

1. Species Identification: Probability scores across 28 bird species.
2. Temporal Localization: Identification of when a bird is singing.
3. Frequency Localization: Identification of the frequency range (Hertz) of the vocalization.

📊 Dataset

The data is sourced from the Fully-annotated soundscape recordings from the Southwestern Amazon Basin.

• Audio: 21 hours of 48 kHz recordings (resampled to 32 kHz).
• Labels: 14,798 manually annotated time-frequency bounding boxes.
• Scope: Focused on the 28 most frequent species (those with >150 occurrences) to mitigate class imbalance.

🏗 Project Structure

├───src
│   │   data_review.ipynb       # Exploratory Data Analysis
│   │   train_iso.ipynb         # Isolated training experiments
│   ├───classes
│   │       BirdDataset.py      # Custom PyTorch Dataset (Log-Mel conversion)
│   │       BirdModels.py       # CRNN Architecture (CNN + GRU)
│   ├───common
│   │       metrics.py          # Macro-Averaged F1, Dice, and Precision/Recall
│   │       plotting.py         # Result visualization
│   └───scripts
│           train_config.json   # Hyperparameters
│           train_script.py     # Training pipeline in script form
└───stable_logs                 # Saved metrics and training plots

🚀 Model Architecture

The model utilizes a CNN backbone for spatial feature extraction from spectrograms, followed by Gated Recurrent Units (GRU) to capture the temporal rhythm of bird songs.

• Primary Metric: Macro-Averaged F1 Score.
• Loss Function: BCEWithLogitsLoss with positive weights for sparse labels.

📈 Results & Detections

The model demonstrates relative precision in identifying most recurring species in various environments with multiple overlapping vocalizations.

Training Progress

To improve the signal-to-noise ratio, the training dataset was augmented by removing background noise in Adobe Audition.

Because bird vocalizations are sparse, all classes were boosted during training using pos_weight in the BCEWithLogitsLoss function.

Below are examples of the model's ability to detect species like the Thrush-like Wren and Amazonian Grosbeak.