A Comparative Analysis of SVM Approaches for Binary Snore Detection Using MFCC Features

Snoring may indicate serious sleep disorders, making accurate detection and classification of snore sounds crucial for early diagnosis and intervention. While existing high-performance algorithms are available, they are mostly based on complex deep learning models that often lack interpretability and explainability, and risk overfitting with insufficient data. We propose a simpler approach using support vector machine (SVM) with Mel-frequency cepstral coefficients (MFCCs) as frequency-based features, offering greater potential for explainability. We evaluated linear, quadratic, cubic, and Gaussian SVM models using two datasets from Kaggle: a Snoring dataset with 1-second duration signals and a subset of ESC-50 dataset (called ESC-10) with 5-second duration signals. We performed 5-fold cross-validation on three datasets: Snoring, ESC-10, and their combination. Nonlinear SVM models with MFCC features achieved performance metrics (except sensitivity, precision, and F1 score for the ESC-10 dataset) ranging from 93.2% to 98.4%, with accuracies between 93.5% and 98.0%. For the ESC-10 dataset, sensitivity, precision, and F1 scores were notably lower due to strong class imbalance. These performances are comparable to existing complex deep learning models, demonstrating that the SVM approach effectively detects snoring sounds with simpler models and fewer parameters. While the combined dataset results indicate the feasibility of handling different signal durations using the same MFCC format, cross-dataset testing revealed significant generalization challenges. High-quality datasets are necessary to better understand these challenges and develop practical high-performance classification algorithms for snore detection.