Construction and Comparison of Snoring Detection Models Based on Machine Recognition

In the field of sleep health monitoring, the accuracy and efficiency of snoring identification, as a key indicator for assessing sleep disorders, are crucial for the early detection and intervention of health problems such as obstructive sleep apnea syndrome (OSA). However, traditional manual auscultation methods are limited by subjective judgment and resource constraints, which make it difficult to meet the demands of largescale screening and real-time monitoring. In this paper, we use hybrid features, fused from Mel Frequency Cepstral Coefficients (MFCC) and Short-Time Fourier Transform (STFT), as inputs. We then comparatively analyze the performance of machine learning algorithms, including Decision Tree, Random Forest, Logistic Regression, K-Nearest Neighbors (KNN), and XGBoost, as well as deep learning models such as Deep Neural Networks (DNNs), Long Short-Term Memory Networks (LSTMs), and Transformer models. The experimental results demonstrate that the KNN algorithm achieves the best accuracy and performance in the snoring sound recognition task under the same feature processing conditions. This study not only verifies the effectiveness of hybrid features for snoring recognition but also highlights the superiority of the KNN algorithm, offering an efficient and reliable solution for snoring monitoring and diagnosis. Additionally, it provides theoretical support and technical references for the development of portable and intelligent sleep monitoring devices, contributing to improved accuracy and widespread adoption.