A Study on Single-channel Non-stationary Noise Suppression for Cardiac Sound Takashi SUDO *† , Hirokazu TANAKA *‡ , Chika SUGIMOTO ‡ and Ryuji KOHNO ‡ * Toshiba Corporation, Minato-ku, Tokyo, Japan † Graduate School of Engineering, Yokohama National University, Hodogaya-ku, Yokohama, Japan ‡ Faculty of Engineering, Yokohama National University, Hodogaya-ku, Yokohama, Japan Abstract—In recent years, the need for health care continues to grow, thereby leading to the widespread use of home-oriented health care and medical equipment. An example of the same is the use of the digital stethoscope. Our research aims to accurately monitor a user’s cardiac status 24 hours a day at any location using a wearable device with an in-built digital stethoscope. A presently existing challenge is that the cardiac sound monitored with digital stethoscopes is often mixed with ambient noise or overlapped with vibrations from body movement. This study investigated a means to suppress sudden non-stationary noise from a single-channel stethoscope signal. We propose a method in which we assume the excitation signal to be a non-stationary noise component in the linear prediction analysis. Compared with the spectral subtraction method, which continually updates the noise spectrum, our method improves the spectrum distortion by 2.0 dB, and improves the degree of noise suppression by 4.6 dB. I. I NTRODUCTION In recent years, the need for health care continues to grow [1], thereby leading to the widespread use of home-oriented health care and medical equipments such as the body fat meter, the sphygmomanometer, and the pulse meter [2]. One such ordinary home-oriented medical product is the stethoscope, and there is ongoing research and development on the digital stethoscope [2]. Since the stethoscope is inexpensive and can be used easily, as compared with an electrocardiograph (ECG) for example, it has been used for many years for the auscultation of cardiac and lung sounds during medical examinations. Cardiac sounds result from motion of heart valves, and the stethoscope obtains useful information regard- ing the first heart sound (S1), the second heart sound (S2), anomalous sounds and cardiac murmurs by which we can judge a hearts condition [3], [4]. Also, illnesses can be detected by auscultation of a cardiac murmur — information that cannot be obtained by an ECG. Examples include atrial fibrillation, aortic valve stenosis, aortic regurgitation, and mitral stenosis [4]. The cardiac sound is processed as a digital signal, and research applied to cardiac disease diagnosis is advancing [5]– [12]. Recent research objectives are to monitor health and life conditions 24 hours a day and at any location using small wearable devices through which data is digitally acquired and health services are subsequently provided [13], [14]. In this study, we aim to acquire a user’s cardiac sounds using a wearable device with an in-built digital stethoscope that has direct contact with skin near the heart. When acquiring cardiac sound using a digital stethoscope, the monitored sound is often mixed with ambient noise or overlapped by vibrations from body movement. In many cases, these interferences are sudden non-stationary noises. Therefore, although noise sup- pression processing is needed, only noise should be suppressed without the degradation of any important cardiac signals. Techniques are not yet established for suppressing sudden non-stationary noise in single-channel input signals in digital acoustic signal processing. However, various methods are still in the research stage. On the other hand, research on the suppression of ambient environmental noise using digital stethoscopes with two microphones has advanced [15]. In this paper, we propose a noise suppression method for a single-channel digital stethoscope that is inexpensive and has a reduced weight for use in wearable devices. This paper is organized as follows: Section II describes the present conven- tional methods available, Section III describes our proposed method, and Section IV reports our experimental evaluation. In Section V, we draw our conclusions. II. CONVENTIONAL METHOD A. Without Noise Detection As noise suppression method for telephone calls, the spectral subtraction (SS) method [16] and the MMSE-STSA method [17] are known well. These methods detect noise by using characteristic differences in speech and noise in the time or frequency domain. They then estimate the spectrum of noise, set up a suppression gain using this spectrum, and sup- press this noise by multiplying a short-time spectral amplitude by the suppression gain. For stationary noise, these methods have achieved a high amount of noise suppression. With these methods, the amount of non-stationary noise suppression tends to be dependent on the speed with which the noise spectrum is updated. To suppress non-stationary noise, this method suppresses hit sounds, such as noise from a keyboard, which are generated suddenly [18]. This method may be used for limited non- stationary noise, extracting non-stationary noise based on its characteristics, and estimating the spectrum of noise. With this method, the amount of non-stationary noise suppression tends to depend upon the accuracy in extracting this noise. ISMICT 2014 1569882813 1