Audio Event Tagging Using SVM and 2D-CNN-LSTM Models

About The Project

This project aims to classify environmental sounds into predefined categories using both traditional machine learning methods (SVM) and deep learning techniques (2D-CNN-LSTM). The dataset used, ESC-50, comprises 2000 audio recordings across 50 sound classes.

Objectives

• Develop and evaluate SVM and 2D-CNN-LSTM models for audio event tagging.
• Explore the effects of feature extraction techniques on model performance.
• Investigate the impact of data augmentation on classification accuracy.

Key Highlights

• ESC-50 dataset: Balanced with 40 recordings per class across five categories (e.g., animal sounds, human non-speech sounds).
• SVM models trained with various feature extraction methods, such as MFCCs and ZCR.
• Deep learning models incorporating log-mel spectrograms and data augmentation.

Methodology

Dataset

• ESC-50: 2000 audio recordings, each 5 seconds long, categorized into 50 classes.
• ESC-10: A subset of ESC-50 with 10 easily distinguishable classes.

Preprocessing

• Feature Extraction for SVM: MFCCs, ZCR, energy statistics, and their derivatives.
• Log-Mel Spectrograms: Used as input for 2D-CNN-LSTM models.
• Data Augmentation: Noise addition, pitch shifting, and time stretching.

Models

1. SVM:

   • Feature extraction techniques include PCA and random frame selection.
   • Approaches: One-vs-One (OVO) and One-vs-Rest (OVR).

2. 2D-CNN-LSTM:

   • Combines convolutional layers for spatial feature learning with LSTM layers for temporal dependencies.
   • Trained with augmented data using AdamW optimizer and early stopping.

Contributors

• Wong Wei Kang
• Anusha Porwal

License

Distributed under the MIT License. See LICENSE for more information.

Acknowledgments

• Dataset: ESC-50 by Karol J. Piczak
• Libraries: Librosa, Audiomentations, TensorFlow, Scikit-learn