Researchers have built a lightweight AI pipeline that could help more accurately identify upper airway obstruction by converting snoring sounds into time-frequency images, but real-world clinical trials remain the next hurdle.
Research: Classification of snoring by deep time-frequency features. Image credit: Kleber Cordeiro / Shutterstock
In the Journal’s recent “Press Articles” scientific reportresearchers proposed a heterogeneous integration framework for snoring source classification.
Snoring is the main symptom associated with obstructive sleep apnea and is caused by obstruction or vibration of upper airway structures such as the epiglottis, base of the tongue, lateral oropharyngeal walls, and soft palate. The anatomical cause of snoring can be identified non-invasively by classifying snoring audio signals. However, current classification methods struggle with limited data, insufficient time-frequency information integration, and unbalanced class distribution.
About research
In this study, researchers proposed a heterogeneous integration framework for snoring source classification. Its framework includes three core modules: short-time Fourier transform (STFT)-based spectrogram generation, pretrained convolutional neural network (CNN) feature extraction, and support vector machine (SVM) classification. Therefore, STFT generates a spectrogram by transforming the snoring audio signal while preserving time-frequency information.
Additionally, advanced time-frequency features are extracted from the spectrogram using a pre-trained CNN. Finally, we train an SVM classifier based on the extracted features to classify the snoring audio signal into four categories. The team tested their model on the Munich-Passau Snoring Sound Corpus (MPSSC). This corpus bundles snoring recordings from the soft palate, tongue base, epiglottis, and oropharyngeal lateral wall, categorized as V, T, E, and O, respectively.
The team split the MPSSC dataset into training, development, and test sets. These snoring class audio samples were unevenly distributed in the training set, with class V samples accounting for 56.9% and class E samples accounting for 10.7%. Therefore, an upsampling technique was adopted to make the sample number more uniform. The team then applied a STFT with a 512-sample window at a sampling rate of 44.1 kHz to generate the spectrogram.
Spectrograms were resized to meet the input requirements of two pre-trained CNNs, VGG19 and AlexNet, and 4096-dimensional features were extracted from fully connected layers 6 (fc6) and 7 (fc7). Additionally, an L2 regularization SVM was trained based on the extracted features. AlexNet fc7 with Viridis color mapping was found to be the best performing combination, yielding unweighted average recall (UAR) of 46.0% and 67.1% on the development and test sets, respectively.
To assess the contribution of each component within the framework, we performed ablation analysis by modifying or excluding individual modules without changing other conditions. Removal of STFT and replacement of the spectrogram by a waveform-based image representation reduced the UAR to 54.3%, a decrease of 12.8 percentage points. This indicates that explicit time-frequency information is important for feature extraction.
Additionally, replacing the SVM with a fine-tuned fully connected layer reduced performance by 7.5 percent, highlighting the importance of classifier selection. In particular, replacing the pre-trained CNN with handcrafted features reduced the UAR by 21.3 percentage points. Next, the team evaluated the effectiveness of the proposed framework against various traditional methods.
These include Mel-frequency cepstral coefficients with SVM (MFCC + SVM), end-to-end CNN, CNN long short-term memory (LSTM) baseline, dual convolution gated recurrent unit (DualConvGRU), audio spectrogram transformer (AST), WavLM, and wav2vec 2.0. These methods were evaluated on the same training, development, and test sets of MPSSC.
Most techniques showed higher UAR on the test set than the development set, with the proposed framework achieving the largest improvement (21.1 percentage points). Furthermore, the proposed framework exhibits better performance than MFCC+SVM, indicating that the handcrafted acoustic features are unable to capture the complex patterns of the snoring audio signal. We also reported a slightly higher test set UAR than end-to-end CNN.
Although DualConvGRU showed a higher UAR than the proposed framework on the development set, the improvement from the development set to the test set was only 8.8 percentage points. The proposed model also reported higher test set UAR than advanced audio models such as WavLM, AST, and wav2vec 2.0. Finally, in the confusion matrix analysis, the DualConvGRU model showed significant confusion between the V and O classes, while the proposed framework achieved a more balanced recall profile, while O was still difficult and T recall decreased.
conclusion
In summary, this study described a snoring classification model based on STFT spectrogram, pretrained CNN feature extraction, and SVM classification. We achieved a UAR of 67.1% on the MPSSC test set. This is the highest reported value among the compared methods. Removing a single module reduced the UAR by 7.5 to 21.3 percentage points, highlighting the complementary role of the modules. Further research is needed to independently validate this model on external clinical datasets and improve its generalizability and robustness.
