Audio Classification (ESC-50) - SVM & Random Forest

A classical machine learning pipeline for environmental sound classification, built on the ESC-50 dataset. The project covers manual feature extraction, feature analysis, model training, hyperparameter tuning, and evaluation using SVM and Random Forest classifiers.

Best model accuracy: 86% - surpassing human-level performance (~81%) on ESC-50

---

Overview

The ESC-50 dataset is an industry-standard benchmark for environmental sound classification. This project focuses on 10 selected classes of diverse sound categories, using manually extracted acoustic features to train and tune classical ML classifiers.

The supervised learning component was developed independently as part of the Applied Machine Learning course at Hanze University of Applied Sciences (2024–2025).

Selected Classes

Category	Classes
Animals	Chirping birds, Frog, Insects
Natural	Thunderstorm
Urban	Airplane, Car horn, Vacuum cleaner
Human	Breathing, Brushing teeth
Mechanical	Clock tick

---

Pipeline

1. Import required libraries and modules
2. Read audio file metadata
3. Select 10 classes (2 from each major ESC-50 category)
4. Split data using predefined folds (to prevent data leakage)
5. Visualize waveforms and spectrograms (1 instance per class)
6. Extract features: MFCC, RMS, Zero Crossing Rate, Spectral Centroid, Spectral Roll-off, Spectral Contrast
7. Aggregate features using mean and standard deviation
8. Visualize features using pair plots
9. Analyze correlation coefficients to ensure feature decorrelation
10. Create feature subsets for comparative experiments
11. Find initial hyperparameter ranges using GridSearchCV
12. Fine-tune hyperparameters on selected ranges
13. Select best model per feature set
14. Evaluate using Confusion Matrix and ROC-AUC curves
15. Train and compare Random Forest classifier

---

Results

Model	Feature Set	Test Accuracy
SVM (Linear, C=0.1)	All features	86%
SVM	Feature Set 1	~65%
SVM	Feature Set 2	~74%
SVM	PCA-reduced	~37%
Random Forest (1000 trees)	All features	Comparable to SVM

ROC-AUC (best SVM): 0.97–0.98

Human-level performance on ESC-50 is reported at ~81%. The best model in this project surpasses that benchmark using only classical ML and manually extracted features.

---

Key Techniques

• Feature extraction: MFCC, ZCR, RMS, Spectral Centroid, Roll-off, Contrast via Librosa
• Cross-validation: PredefinedSplit (respects ESC-50 fold structure to avoid data leakage)
• Hyperparameter tuning: Two-stage, GridSearchCV and manual (broad → narrow)
• Evaluation: Confusion matrix, per-class ROC-AUC, macro-average F1

---

Project Structure

audio-classification-esc50/
│
├── audio_classification_esc50.ipynb   # Full pipeline notebook
├── requirements.txt                   # Dependencies
└── README.md

Note: The ESC-50 dataset is not included. Download it from the official repository.

---

How to Run

1. Clone the repository:

git clone https://github.com/MostafaJahanian/audio-classification-esc50.git
cd audio-classification-esc50

2. Install dependencies:

pip install -r requirements.txt

3. Download the ESC-50 dataset and update the file paths in the notebook accordingly.

4. Run the notebook top to bottom.

---

Author

Mostafa Jahanian LinkedIn · GitHub