A Bag of Wavelet Features for Snore Sound Classification

Area of Science:

• Biomedical Engineering
• Otolaryngology
• Computer Science

Background:

• Snore sound (SnS) classification aids in diagnosing sleep-related breathing disorders.
• Identifying the precise location of upper airway obstruction and vibration is crucial for targeted surgical interventions.
• Machine listening methods offer a non-invasive approach to analyze complex acoustic signals like snoring.

Purpose of the Study:

• To develop an enhanced machine listening method for accurate snore sound classification.
• To improve the identification of obstruction and vibration locations within the upper airway using acoustic features.
• To evaluate the performance of a novel bag-of-audio-words approach combined with wavelet features.

Main Methods:

• Extraction of low-level wavelet features from snore sound data.
• Application of a bag-of-audio-words model to enhance wavelet features.
• Utilizing a Naïve Bayes classifier for snore sound classification.
• Validation using a dataset of 219 subjects undergoing drug-induced sleep endoscopy.

Main Results:

• The proposed method achieved an unweighted average recall of 69.4%.
• This significantly outperformed the official baseline (58.5%) and the INTERSPEECH COMPARE Challenge 2017 winner (64.2%).
• The enhanced wavelet features demonstrated superior performance compared to conventional acoustic features.

Conclusions:

• The developed bag-of-audio-words approach effectively enhances snore sound classification.
• This method shows significant potential for improving the diagnosis and surgical planning of sleep-related breathing disorders.
• The findings highlight the efficacy of advanced machine listening techniques in analyzing complex biomedical acoustic signals.