ESC50 Challenge

Project Description

This project focuses on the ESC50 Challenge. The ESC-50 dataset is a labeled collection of 2000 environmental audio recordings suitable for benchmarking methods of environmental sound classification. The dataset consists of 5-second-long recordings organized into 50 semantic classes (with 40 examples per class), loosely arranged into 5 major categories.

Key Components

1. Data Generation: The dataset_ESC50.py file is used to generate the data.
2. Training Pipeline: The Train_crossval.py file provides the training pipeline, in which a 5-fold cross-validation training is conducted.
3. Model Storage: The trained models are stored in the result folder.
4. Testing: The test_crossval.py file tests all 5 folds and calculates a mean accuracy.
5. Results: The average accuracy of the model is 83.2%.

Tech Stack

• Torch
• scikit-learn (sklearn)
• librosa