Snoring classification with deep time-frequency features

Snoring is a primary symptom of Obstructive Sleep Apnea (OSA), and its accurate classification is crucial for non-invasively locating upper airway obstructions. However, existing methods struggle with insufficient data, imbalanced classes, and inadequate integration of time-frequency information. To address these challenges, this paper proposes a heterogeneous integration framework combining Short-Time Fourier Transform (STFT), pre-trained CNNs, and an L2-regularized Support Vector Machine (SVM). First, STFT is employed to convert snore signals into spectrograms with a perceptually uniform Viridis colormap, preserving critical time-frequency structures. Second, deep time-frequency features are extracted from the fc7 layer of a pre-trained AlexNet, which inherently mitigates the problem of limited labeled data. Finally, an L2-regularized SVM replaces the standard softmax classifier to counteract overfitting under high-dimensional, small-sample conditions. Experiments on the Munich-Passau Snore Sound Corpus demonstrate that our method achieves a test set Unweighted Average Recall of 67.1%, outperforming state of the art methods including end to end Convolutional Neural Networks and Transformer based audio models. Ablation studies confirm that removing any single component including Short-Time Fourier Transform, pre-trained Convolutional Neural Network, or Support Vector Machine causes a significant performance drop of up to 21.3%. The proposed framework provides an effective, generalizable, and data efficient solution for snore source localization.