TL;DR
This paper introduces a novel dual source-filter model, inspired by speech production, to classify different types of snorers based on their audio. By modeling the snore generation process and extracting features from this model, they achieve improved classification accuracy compared to existing methods, offering a more precise way to identify the location of airway obstruction.
What Problem Does This Paper Address?
Snoring is a common symptom of serious conditions like obstructive sleep apnea (OSA). Identifying the specific location of airway obstruction (e.g., velum, oropharyngeal walls, tongue, epiglottis) is crucial for effective treatment. Current methods for classifying these obstruction locations often rely on expensive and time-consuming procedures like drug-induced sleep endoscopy (DISE). While acoustic analysis of snore sounds has shown promise, existing approaches primarily focus on extracting well-known acoustic features without modeling the underlying generative process of the snore sound itself. This paper aims to bridge this gap by proposing a model that better represents how snore sounds are produced, leading to more accurate classification.
Methods
The core of this paper is the proposal of a dual source-filter model for snore audio generation. This model is analogous to the source-filter model used in speech production. It posits that snore sounds are generated by two acoustic tubes and two excitation sources:
- First Acoustic Tube (F1): Modeled as an all-pole filter, representing the airway from the lungs to the point of obstruction. This tube is excited by white noise, simulating airflow from the lungs.
- Second Acoustic Tube (F2): Modeled as another all-pole filter (though a pole-zero filter is suggested as a future improvement), representing the airway from the obstruction point to the lips/nose. This tube is excited by a combination of the output from the first filter and an impulse train, which captures the vibrations at the obstruction site.
The model parameters (filter coefficients) are estimated using a novel method that involves:
- Spectral tilt estimation and removal: To preprocess the snore audio.
- Snorer Beat Closing Instant (SBCI) estimation: Identifying key moments in the snore cycle, analogous to glottal closure instants in speech.
- Window Construction: Creating specific windows to attenuate certain parts of the snore beat cycle, focusing on the influence of the obstruction.
- Weighted Linear Prediction (WLP): Estimating the coefficients of the first filter (F1).
- Inverse Filtering: Removing the effect of the first filter to isolate the second filter's characteristics.
- Sparse Estimation: Using an L1-norm constraint to automatically infer the order of the first filter, making the model more robust.
Features are then extracted from these estimated model parameters. Finally, a Support Vector Machine (SVM) classifier is used to categorize the snorer into one of the four groups (VOTE).
Datasets
The experiments were conducted on the MUNICH-PASSAU SNORE SOUND CORPUS (MPSSC). The paper notes that the dataset is imbalanced, with a higher proportion of V and O group samples, reflecting their prevalence in normal sleep. The development set of 283 snore audios was used for evaluation.
Key Insights and Ideas
- Snore Generation Mimics Speech: The paper successfully demonstrates that a source-filter model, a cornerstone of speech production modeling, can also effectively approximate the generative process of snore sounds.
- Dual Nature of Snore Excitation: The idea of two distinct excitation sources – white noise from the lungs and quasi-periodic vibrations at the obstruction – is a key insight into snore acoustics.
- SBCIs as Crucial Landmarks: Identifying SBCIs is vital for accurately estimating model parameters, highlighting their importance in characterizing snore dynamics.
- Sparsity for Model Order Inference: Employing sparsity constraints to automatically determine the order of the first filter is an elegant way to handle the unknown obstruction location without requiring manual feature engineering for filter order.
- Filter Coefficients as Discriminative Features: The proposed model's filter coefficients, particularly when combined, prove to be highly discriminative for classifying snorer groups, outperforming traditional acoustic features.
Key Results
The proposed dual source-filter model achieved an unweighted average recall (UAR) of 49.58% on the development set. This represents a significant improvement of approximately 3% (absolute) over the INTERSPEECH-2017 snoring sub-challenge baseline technique.
Specifically, the coverall feature (derived from the combined filter coefficients) with an RBF kernel achieved the highest UAR. The paper also highlights that the proposed features showed substantial improvements in classifying the T and E groups (e.g., ~41% and ~27% improvement for T over CF) compared to baseline methods, especially on imbalanced datasets.
Limitations and Future Work
The authors acknowledge several limitations:
- Independence of Filters: The current model treats the two filters independently, whereas there might be interactions between them.
- All-Pole Filter Assumption: The second filter is modeled as an all-pole filter, which is an approximation. A pole-zero filter might offer a more accurate representation.
Future work suggested includes:
- Modeling Filter Interaction: Incorporating the interaction between the two filters into the model.
- Pole-Zero Filter: Implementing a pole-zero filter for the second acoustic tube.
- Improving SBCI Detection: Enhancing the SBCI detection method, especially for noisy snore audio with lower fundamental frequencies, to further improve classification accuracy.
Summary updated