Audio Source Separation for Reducing Sleeping Partner Sounds: Simulation Study

Millions of people worldwide experience sleep disorders and untreated sleep problems can lead to poor health, accidents, and even death. Currently, for increasing the accessibility of diagnosis, the field of sleep disorders evaluation is going towards home-based testing by using non-contact sensors, such as microphones. However, when recording a subject in an at-home environment, the breathing recording might be distorted by sounds such as TV, air conditioner and sleeping partner. In this study, we implement source separation algorithms in order to separate the desired subject from their sleeping partner. To do that, we developed a PC simulation: two sleeping partners (recorded separately) were mixed with respect to their propagation delays, distances from the microphones and room reverberations. These mixtures went through several separation algorithms: principal component analysis (PCA), fast independent component analysis (ICA) and kurtosis ICA that relies on the maximization of signal statistic, and degenerate un-mixing estimation technique (DUET), that rely on time-frequency representation of the signal. We evaluated the separation results in different SNRs and reverberation times. The results showed that even in low SNR (-20dB) we obtain good separation (MSE=1.29*10^-3) especially by ICA and PCA, while in high SNR (>0dB) the DUET performed better (MSE=0.99*10^-3). Moreover, whereas without reverberations the separation results were good, in the reverberated mixtures the quality of separation depended on the room characteristics, such the absorption coefficient, room dimension, microphone distance etc. The separated desired source signal can be further analyzed by systems for snoring evaluation, sleep apnea, and macro-sleep stage estimation.