On the Extraction of the Snore Acoustic Signal by Independent Component Analysis F. Vrins 1 , J. Deswert, D. Bouvy, V. Bouillon, J. A. Lee 1 , C. Eug` ene 2 and M. Verleysen 1 1 Microelectronics Laboratory, Machine Learning Group 2 Electrotechnics and Instrumentation Laboratory Department of Electricity Universit´ e catholique de Louvain (UCL), Place du Levant 3, 1348 Louvain-la-Neuve, Belgium {vrins, lee, verleysen}@dice.ucl.ac.be, eugene@lei.ucl.ac.be Abstract Physicians are interested in the acoustic signal of snore, because it allows them to diagnose the patient and eventu- ally to avoid several dangerous accidents. Today, its mea- sure is not satisfactory for various reasons. In this pa- per, we explore a new way to measure this signal: Blind Source Separation (BSS). We give encouraging results of source separation in this application but also stress the obstacles which prevent a perfect separation when BSS is achieved by the classic linear, instantaneous and un- noiseless model of Independent Component Analysis, an undoubtfully very promising signal processing method in the biomedical world. Key Words Snoring, non-invasive measurement, independent compo- nent analysis. 1 Introduction Recently, new signal processing techniques appeared in the biomedical application’s world (a.o. electroencephalo- gram [1], fetal-electrocardiogram [2, 3], . . . ), like Blind Sources Separation (BSS). BSS consists in recovering sta- tistically independent sources by analyzing only mixtures of them (sources are -supposed to be- unknown). Un- der certain assumptions, Independent Component Analy- sis (ICA) is able to achieve BSS. However, several ‘real- world’ problems prevent ICA to recover perfectly each source. Indeed, the theoretical model does not exactly cor- respond to reality. Nevertheless, in many cases, ICA could give interesting results to the physicians. A new biomedical application is considered here: the extraction of an acoustic snore signal. This is an impor- tant application: snoring has many consequences on the patient’s life, and the acoustic signal could really help the physicians to diagnose breathing problems and to evaluate the risk incurred by the patient [4]. Today, it is difficult to record this signal. A brief analysis of the situation shows that this problem corresponds to the framework of BSS, and an attempt of solution by ICA seems natural. In this paper, we explore a first tentative to extract the acoustic snore sig- nal by Independent Component Analysis, and we stress the obstacles of the BSS of acoustic signal measured in real environment in general which make the problem complex. The paper is organized as follows: section 2 describes the snoring phenomenon, and stresses its interest. Sec- tion 3 presents the framework of BSS. Considerations are given in section 4 about the feasibility of the snore sig- nal extraction by ICA. We present the signal processing in order to achieve the extraction of the snore signal in section 5. We show results for mixtures including snore, narrow-spectrum and wide-spectrum signals in section 6. Discussions and conclusions are given in section 7 and 8. 2 The snoring The snoring is a well-known physical phenomenon, ap- pearing (most often on male patients) during sleep. The acoustic signal of snore is caused by the vibrations of the pharynx slack tissues. Several snore levels exist, classified with respect to their power spectral content, position of the patient during snoring, etc. In addition to the acoustic nui- sances, snoring can cause others physical problems for peo- ple which are directly suffering from, like tiredness during the day. It also increases the probability of cardiovascular and cerebral accidents. 2.1 Specificity of the signal The acoustic signal of snore allows diagnosing the patient. Its measure could be associated with other physical data, like electroencephalograms, electro-oculograms, etc. The power is concentrated under 5kHz. The time structure of a typical snore signal is shown in Figure 1 (f samp , the sample frequency, is 44.1 kHz). The snore signal is almost periodic and stationary on windows larger than its period. 2.2 Measurement: state-of-the-art Currently, two methods exist to record the snore signals: microphones and piezoelectric sensors. Among the micro- 417-049 326