Researchers announced on July 5 that they have developed an artificial intelligence pipeline capable of identifying the anatomical origin of snoring by analyzing audio signals. The study, published as an 'Article in Press' in Scientific Reports, describes a framework that integrates short-time Fourier transform-based spectrogram generation, pretrained convolutional neural network feature extraction, and support vector machine classification.
Snoring is a primary symptom associated with obstructive sleep apnea and results from obstructions or vibrations in upper airway structures such as the epiglottis, tongue base, lateral oropharyngeal walls, and soft palate. Current methods for classifying the source of snoring face challenges including limited data and poor integration of time-frequency information.
The researchers tested their model using the Munich-Passau Snore Sound Corpus (MPSSC), which contains labeled recordings from different parts of the upper airway. To address uneven distribution among classes within their training data—where one class made up 56.9% and another only 10.7%—they used upsampling techniques to balance sample counts before generating spectrograms with a 512-sample window at a sampling rate of 44.1 kHz.
Spectrograms were resized for compatibility with two pretrained convolutional neural networks: VGG19 and AlexNet. Features extracted from fully connected layers were then classified using an L2-regularized support vector machine. The combination of AlexNet's fc7 layer features with Viridis color mapping yielded the best performance: an unweighted average recall (UAR) of 67.1% on test data.
Ablation analyses showed that removing or modifying any core module reduced performance significantly; omitting explicit time-frequency information dropped UAR by nearly thirteen percentage points, while replacing CNN features with handcrafted ones led to a decline exceeding twenty percentage points. When compared against other conventional methods—including Mel frequency cepstral coefficients plus SVM, end-to-end CNN models, DualConvGRU networks, audio spectrogram transformers (AST), WavLM, and wav2vec 2.0—the proposed approach demonstrated higher test-set UAR than all except DualConvGRU on development data but achieved greater improvement between development and test sets overall.
The authors concluded that their heterogeneous integration framework offers improved accuracy over existing approaches for non-invasively identifying where snoring originates within the airway based on sound recordings alone, but noted further validation is needed using external clinical datasets.
