Snore-GANs: Improving Automatic Snore Sound Classification With Synthesized Data.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Automatic snore sound classification (ASSC) is crucial for diagnosing sleep disorders.
• A significant challenge in ASSC development is the scarcity of supervised training data.
• Existing data augmentation methods may not adequately address the complexity of snore sound data.

Purpose of the Study:

• To propose a novel data augmentation method for ASSC using semi-supervised conditional generative adversarial networks (scGANs).
• To address the limitation of insufficient supervised training data in automatic snore sound classification.
• To enhance the diversity and realism of synthesized snore sound data.

Main Methods:

• Developed a semi-supervised conditional generative adversarial network (scGAN) to synthesize realistic high-dimensional snore sound data.
• Implemented an ensemble strategy to mitigate the mode collapse problem inherent in Generative Adversarial Networks (GANs).
• Utilized a widely adopted Munich-Passau snore sound corpus for experimental validation.

Main Results:

• The scGAN-based data augmentation approach significantly outperformed traditional data augmentation techniques.
• The proposed method demonstrated competitive performance compared to other state-of-the-art systems for ASSC.
• Synthesized data required no additional manual annotation, streamlining the data preparation process.

Conclusions:

• Semi-supervised conditional generative adversarial networks (scGANs) offer a powerful solution for data augmentation in automatic snore sound classification.
• The scGAN approach effectively addresses the data scarcity issue, leading to improved ASSC performance.
• The ensemble strategy enhances the robustness and diversity of generated data, making it suitable for training robust ASSC models.