TL;DR
This paper explores the novel idea of identifying individuals based on their snoring sounds, aiming to enable more practical and less intrusive sleep monitoring. By applying machine learning models like GMM and DNN to snore recordings, they demonstrate that snoring patterns contain unique signatures, achieving over 90% accuracy in both identifying and verifying users.
What Problem Does This Paper Address?
Traditional methods for monitoring sleep disorders like Obstructive Sleep Apnea-Hypopnea Syndrome (OSAH) often rely on polysomnography (PSG). PSG is effective but inconvenient, requiring overnight stays in sleep clinics and extensive equipment. While smartphone-based monitoring offers a less invasive alternative, a significant challenge arises in shared sleeping environments. When multiple people share a room, raw audio recordings can capture the snores of bed partners, leading to false alarms and misdiagnosis. This paper tackles the problem of distinguishing between different individuals' snores in such shared environments, proposing snore-based user recognition as a solution.
Methods
The researchers investigated three machine learning models for snore recognition:
- Gaussian Mixture Model (GMM): A standard statistical model used for signal approximation and speaker recognition. They explored GMMs with varying numbers of components.
- GMM-Universal Background Model (GMM-UBM): An extension of GMM that leverages a pre-trained background model from all users to improve performance, especially when individual enrollment data is limited. This helps reduce dependence on subject-specific data.
- Deep Neural Network (DNN): Inspired by its success in speaker and cough recognition, a DNN was employed to extract "snore embeddings." This involved a multi-stage process: stacking MFCC features with adjacent frames as input, training a DNN to output a one-hot identity vector, and then using the frozen layers to generate utterance-level embeddings. These embeddings were then averaged to create a subject's snore embedding.
For all models, audio signals were preprocessed using Short-Time Fourier Transform (STFT) and Mel-Frequency Cepstral Coefficients (MFCCs) were extracted as features.
Datasets
The study utilized the MPSSC (Munich-Passau Snore Sound Corpus) dataset. This corpus contains snore sounds collected from Drug-Induced Sleep Endoscopy (DISE) examinations, with a focus on clear, single-vibration events. The dataset comprises 828 utterances from 219 subjects, recorded at a 16KHz sampling rate. The authors selected subjects with at least five utterances for their experiments.
Key Insights and Ideas
- Snoring as a Biometric: The core hypothesis is that snoring sounds, like speech or fingerprints, possess unique characteristics that can identify an individual. This paper provides strong evidence supporting this idea.
- DNN Embeddings Outperform Traditional Models: The study highlights that DNN-generated snore embeddings offer superior clustering and discriminative power compared to raw MFCC features, visually demonstrated through t-SNE plots.
- Practicality in Shared Environments: The ability to distinguish snores is crucial for making passive, smartphone-based sleep monitoring truly useful and reliable in multi-person households or shared clinical spaces.
Key Results
- Identification Accuracy: All three models achieved identification accuracies exceeding 90%.
- GMM-UBM Superiority: The GMM-UBM model consistently outperformed the standard GMM, showing improved identification accuracy across different component numbers, ranging from 95% to 100%.
- DNN Verification Performance: The DNN-based model demonstrated the best verification performance, achieving over 95% sensitivity and specificity. ROC curves also indicated the DNN model's superior performance.
- Efficiency: Promising results were achieved with as few as four snoring samples per user for enrollment.
Limitations and Future Work
- Limited Enrollment Data: While the study achieved good results with limited enrollment data (four utterances), the performance of GMM-based models slightly decreased with more components, suggesting a sensitivity to the small training set size.
- Future Work: The authors suggest that this work is a foundational step towards understanding the practicality and feasibility of user recognition via snoring. Future directions could involve exploring more advanced deep learning architectures, larger and more diverse datasets, and real-world deployment scenarios to further refine and validate these systems for ubiquitous snore monitoring.
Summary updated