PRZEGLĄD ELEKTROTECHNICZNY, ISSN 0033-2097, R. 91 NR 2/2015 79 Edward PÓŁROLNICZAK 1 , Michał KRAMARCZYK 2 1,2 West Pomeranian University of Technology, Szczecin, Al. Piastów 17, 70-310 Szczecin doi:10.15199/48.2015.02.20 Computer analysis of the noise component in the singing voice for assessing the quality of singing Abstract. The article focuses on the analysis of a noise in singing voices using samples obtained from choral singers. Analysis of the singing voice quality is a complex task. There are various points of view and methods of analysis used to solve that problem. Musicians consider the voice of singers from the point of view of voice emission. Doctors analyzing the voice of a singer trying to determine his medical parameters. To analyze the quality of singing voice the equipment and tools characteristic for clinical practice can be used. In the analysis of the signals mentioned in the article some methods for the voice analysis have been adopted. The quality parameters of the singing voice were calculated on the basis of recorded samples. This article presents the results of research at the frontier of computer signal analysis and medicine. To achieve the goal in the study a method for the analysis of noise Noise Parameter based on GNE (Glottal-to-Noise Excitation Ratio) was used. Streszczenie. Niniejszy artykuł skupia się na analizie szumu w głosie śpiewaczym. Analiza jakości głosu śpiewaczego jest złożonym zadaniem. W praktyce istnieją różne podejścia do analizy zagadnienia szumu w głosie. Muzycy patrzą na głos śpiewaczy z perspektywy emisji głosu. Lekarze analizując głos śpiewaka próbują ustalić jego parametry medyczne. Stosowane są przy tym narzędzia wykorzystywane do analiz medycznych. W przedstawionej w tym artykule analizie parametry jakościowe głosu śpiewaczego obliczono dla próbek dźwiękowych pozyskanych od śpiewaków chóralnych. Przedstawiono tutaj wyniki badań z pogranicza komputerowej analizy sygnału i medycyny. Aby osiągnąć założony cel wykorzystano metody analizy parametrów szumu na podstawie współczynnika GNE (Glottal-to-Noise Excitation Ratio). Analiza szumu w głosie śpiewaczym Keywords: signal analysis, singing voice, singing quality Słowa kluczowe: analiza sygnału, głos śpiewaczy, jakość śpiewu Introduction The direct motivation to undertake the research in the field of quality of singing voice is the need to support by computer analysis the education of voice production according to the requirements of choral ensembles. One of the goals of the choir is continuous raising the artistic level. The presented research can support this expectation. The quality of the choir depends largely on the individual voice quality and the skills of individual choir singers. The developed criteria of computer quality voice assessment can be useful in work of voice coach and conductor. Properly described evaluation criteria may allow for correction of the selected parameters can support a self- education of voice production. There are various characteristics of voice which can be assessed by experts (and also by computer algorithms). One of the most obvious and easy to assess by human experts is the intonation. It can give the answer for the question if the singer is singing the melody properly. Another easy to recognize parameter of singing voice is vibrato The concept of vibrato is generally known as the waving of frequency and the volume occurring in the voices of singers. Trainers of voice emission define the vibrato as a periodic, pulsing, and amount of intensity and tone, which is experienced as color impression. In this perspective, the vibrato concept is wider than usual waving of frequency. The third is the timbre phenomenon. Although this factor is very subjective in perception, it can be used as an evaluation criterion of singing. Among the characteristics of the timbre, experts point to the bright, dark, sonorous, matte, clean, harsh, warm or cool timbre and others. Some of these characteristics, such as the matt or rough timbre, are perceived negatively in case of singing. Among the various parameters (attributes) of singing voice there is also a phenomenon connected to a noise component. It can cause the impression of humming, hissing or rustling voice which we named here as the noisy voice as all of the listed characteristics are connected to noise component. From the medical point of view it is connected to a disorder of vibration of the vocal folds of the larynx [1]. It can be result of voice disease or just individual voice characteristic of a singer. Sometimes it is connected to a loss of high frequency components in the signal. The presented research may be significant not only for the quality assessment of singing but also for the health of the singer. Noise component in a singing voice can be a kind of warning signal determining that a singer must undergo a diagnosis. Literature review Measures used to determine the degree of noise in voice signal are important group of features used for assessment of voice quality. Due to the fact that that they are largely limited to the assessment of the accuracy of the harmonic structure the methods are commonly named HNR (Harmonic-to-Noise Ratio). In the literature a number of indicators can be found, but the most important are: 1. HNRYumoto (Yumoto, 1982) – is estimated in the time domain: the aver-age glottal period is calculated after the length of each period is normalized, then each segment is compared with the average (segment). The variance of deviations is the measure of noise [4]. n the modified version HNRQ i (Qi, 1992) segments are normalized in the time domain using the dynamic time warping [5]. 2. NNE – Normalized Noise Energy (Kasuya, 1986) – is the ratio of non-harmonic signal energy to the total energy of signal. It is estimated by location of the minima between the harmonics in the range 1-5 kHz, which carry information about the level of noise [6]. 3. HNR, SPI, TNI (MDVP, 1993) parameters – are based on the same trans-formation: spectral comb filtering, separating the harmonic and non-harmonic energy of the signal. Each measure is the ratio of energy value in specific frequency ranges [7]. 4. CHNR – Cepstral HNR (de Krom, 1993) – it involves the cepstrum to determine the baseline of spectrum. After that the ratio of total energy to the base energy of spectrum is determined [8]. Further modifications rely on better fitting the baseline for the spectrum [9]. Often the measure is calculated according to the algorithm in [3]. The method uses discrimination between harmonic and noise energy in the magnitude spectrum by means of a comb-liftering operation in the cepstrum domain. The authors claim the