Relations among Verbal Attributes Describing Musical Sound Timbre in Czech Language O. Moravec, J. Stepanek Musical Acoustics Research Center, Faculty of Music, Academy of Performing Arts in Prague, Malostranské nám. 13, 118 00 Praha 1, Czech Republic, {moravec, stepanek}@hamu.cz, This study of relations among verbal attributes describing musical sound timbre was based on the observation of opinion and experience of Czech music professionals. A set of experiments was realized without listening to any context of sounds. Verbal attributes were collected from the questionaries filled in by 120 respondents at the first stage of the research. In the next stage pair dissimilarity in timbre was found out for 25 most frequently used attributes. Forty three test respondents originated from five groups of music professionals (players of string, wind and keyboard instruments, composers and conductors, sound designers) took part in this experiment. Latent class approach of multidimensional scaling (CLASCAL) was used for the evaluation of results of this experiment. Resulting perceptual space of verbal attributes of all respondents was compared with perceptual spaces of respondent classes established by CLASCAL and with perceptual spaces of professional groups. Three main dimensions of timbre were specified; at least the two dimensions are common for all classes and groups of respondents. In the last experiment respondents were asked to rate each attribute according to the suitability of its use for the description of timbre, for the description of player s performance quality and for the description of artistic expression. 1 Introduction Multidimensional nature of timbre was undoubtedly accepted by majority of researchers. The main interest in timbre research is the question of existence and the search for common system of perceptual dimensions. The dimensions can be interpreted using acoustic characteristics of sounds [1] or verbal attributes [2] or both. When using verbal attributes, the problem arises how to select representative, suitable and complete set of attributes according to tested sound context timbres. Bismarck's attribute pairs [2] were selected from previous studies and validated by the group of experts before using in listening tests. Spontaneous verbal description of tested sounds collected within experiment was used for the interpretation of perceptual spaces in [3]. All above mentioned experiments represent approach leading from timbre description of specific sound context to possible generalization (bottom-up approach). Stumpf [4] has used semantic analysis of verbal attributes and without any experiment (in contemporary point of view) has obtained three dimensions with attributes: 1. dunkel hell 2. stumpf / weich scharf / rauh 3. voll / breit leer / dünn Attributes can be translated into English: 1. dark bright 2. unpointed / soft / rough 3. / wide empty / An approach to the study of timbre described in this contribution can be named top-down, because it starts with the detection of common experience of musicians which is followed with deeper and more precise qualification. The main aims of the research were: 1. To search for verbal attributes (words and expressions) used by musicians for description of musical sound timbre. 2. To construct common perceptual (semantic, verbal) space, to establish its dimensionality and verbally describe individual dimensions. To ask if all musicians share common perceptual space or what are the criteria dividing them into groups with different timbre perception models. 3. To ask musicians for suitability of verbal attributes for the description of timbre or other aspects of musical performance. To fill these tasks the three sound-context free experiments (without listening to any sound stimuli) were carried out. Czech music professionals took part as respondents in these experiments in which their opinion and experience were observed. 2 Method 2.1 Experiment 1 A questionary survey was applied for collection of verbal attributes used by musicians for the description 1601
Table 2: Questions and scales used in suitability test of selected verbal attributes. No. Table Question 1: Questions and scales used in suitability test of selected Judgment verbal scale attributes. 1 I use this attribute for musical No. Question never occasionally Judgement scale often very often I use this attribute for musical 12 I consider it for musical never occasionally partly often very often unappropriate appropriate very appropriate appropriate I consider it for musical 3 Attribute is suitable for description: partly 2 unappropriate appropriate very rather appropriate rather attack appropriate both - attack or steady state steady state Attribute is suitable for description of: rather 34 - timbre of detached tone rather attack both - attack or steady state unappropriate adequate very steady appropriate state of some instrument - timbre of detached tone 45 - kind (quality) of play of tone of some instrument played on some instrument - kind (quality) of playing 56 - artistic expression connected of tone played on some instrument with musical performance - artistic expression 6 connected with musical performance of timbre [5]. The questionary was composed of two parts. In the first part the respondents described their personal profile - age, instrument(s) they play, music profession, education, etc. The second part of the questionary was divided into three subsections. In the first subsection the respondents wrote down in free order the words and expressions which they use for the description of musical sound timbre, in the second one they wrote down groups of synonyms and in the third one groups of antonyms. The first part of the questionary was used for the description of population sample and for the discrimination of respondents into classes according to selected criteria. The second part as a whole was used for building of individual respondent frequency vocabulary. Common frequency vocabulary and partial frequency vocabularies for respondent groups were then created from individual vocabularies [5]. The second and the third subsections of the second part of the questionary were used for the basic determination of word relations [6]. 2.2 Experiment 2 Sound-context free dissimilarity pair test was performed on 25 verbal attributes most frequently used in questionaries. Respondents were asked to quantify the dissimilarity of sound timbre by virtue of their internal image of timbre of sounds described by attributes in presented pair [7]. Obtained dissimilarity matrices were evaluated using latent class approach applied on the weighted Euclidean model (CLASCAL). Overview of multidimensional scaling models is for example in [8], more precise description of CLASCAL procedure can be found in [9]. Latent class approach was used on a set of all test respondents and also on respondent groups defined by their music profession. Resulting perceptual spaces were compared by the way of optimal fitting of external scales [10] generalized to the case of correlated dimensions (embedding of external scales into perceptual space described in [11]). In this case the external scales were the dimension coordinates of compared group perceptual space. This local comparison of individual dimensions can be more sensitive for partial difference of spaces then global comparison (e.g. Procrustean transformation, see [10]). The interpretation of each perceptual space was based on the search of (nearly) orthogonal system of space objects (attributes); the goal was also to find attributes with (nearly) opposite positions in perceptual space. System of n orthogonal attributes in n-dimensional perceptual space was considered as its successful interpretation. 2.3 Experiment 3 Opinion of musicians on the use and suitability of verbal attributes for the description of timbre in more specific aspects of musical performance were detected. Respondents answered six questions (Table 1) asked consecutively for sixty selected verbal attributes (the first 25 most used attributes from common frequency vocabulary and other attributes selected from vocabulary on the base of results of previous experiments). Judgment scales were quantified for processing of results and then basic statistical characteristics for each question and verbal attribute were calculated and compared: median, mean value, etc. 3 Results 3.1 Experiment 1 Finally 120 filled forms of questionary were acquired. The survey was carried out among students and 1602
professors of the Music Faculty of the Academy of Performing Arts in Prague, soloists and members of several symphonic and chamber orchestra from different regions of the Czech Republic, teachers from musical schools and sound designers from recording studios. These groups overlap in some cases, e. g. professors of the faculty are often also members of orchestras or solo players. More detailed description of respondent distribution is in [5]. In the following analysis the division to only main instrument groups (bow, wind, keyboard) according to instruments the respondents play was used in order to reach sufficient number of respondents in each group. The respondents wrote down 1 964 different verbal attributes in total, 30 attributes from common vocabulary have relative frequency greater then 25 % (attribute was used at least by 30 respondents). Partial vocabularies of groups of bow, wind and keyboard instrument players were compared with common vocabulary. Relative and absolute frequencies of attributes in common vocabulary and attribute ranks in common and partial vocabularies are shown in Table 2. The prominent differences among the whole respondent sample and each group of respondents, supported also by prominent differences among pairs of groups, yields to only a small number of verbal attributes which are significantly more or less frequently used by members of studied groups [5]. Table 2: Thirty the most frequently used verbal attributes from common vocabulary with their absolute (f abs ) and relative (f rel ) frequencies and ranks (including ties) in common and partial (according to respondent group) vocabularies. Attributes from to were used in experiment 2 and 3. Verbal attribute Frequency - all respondents Rank in respondent group Czech original English translation f abs f rel all bow wind keyboard ostrý 94 78.3 1 1 1 1 temný 79 65.8 2 5 2 2 měkký soft 78 65.0 3 2 3 4 jasný 75 62.5 4 3 4.5 4 sametový velvety 61 50.8 5 7 6 4 jemný 58 48.3 6.5 9.5 11.5 6.5 kulatý round 58 48.3 6.5 14.5 4.5 8 tupý unpointed 55 45.8 8 11 8 10 drsný 54 45.0 10 12.5 19.5 6.5 světlý bright 54 45.0 10 16.5 11.5 10 tvrdý hard 54 45.0 10 9.5 8 15.5 sladký sweet 53 44.2 12 4 15 22 plný 51 42.5 13 12.5 8 15.5 hrubý rough 46 38.3 14.5 7 15 37 tmavý dark 46 38.3 14.5 21.5 15 15.5 teplý warm 43 35.8 16 21.5 23.5 15.5 zářivý radiant 42 35.0 17 16.5 45 22 čistý pure 40 33.3 18.5 26 15 19 vřelý hearty 40 33.3 18.5 7 31 78 barevný colored 38 31.7 20.5 19 45 12 zvonivý ringing 38 31.7 20.5 39.5 19.5 10 chladný cool 36 30.0 23.5 14.5 31 61.5 průzračný lucid 36 30.0 23.5 24 65.5 15.5 široký wide 36 30.0 23.5 32 15 37 úzký 36 30.0 23.5 39.5 8 37 kovový metallic 34 28.3 26.5 32 31 22 studený cold 34 28.3 26.5 27.5 31 37 svítivý shining 32 26.7 28 48 19.5 37 zastřený blurred 31 25.8 29 19 23.5 28.5 hladký smooth 30 25.0 30 39.5 45 28.5 1603
3.2 Experiment 2 Forty three of the respondents took part in dissimilarity pair test. The constitution of respondent groups according to their music profession is shown in [12]. Retest performed on a subset of attributes exhibited satisfactory reliability of all respondents. The application of CLASCAL procedure resulted in division of respondents into three classes. All respondents shared common model with threedimensional space of verbal attributes, their classes differed in weights of each dimension. No correlation of respondent classification with their music profession was found, the main criterion for discrimination into classes was the manner the respondent used the dissimilarity evaluation scale (e.g. mean value of dissimilarity), see also [7]. CLASCAL procedure was also applied separately on results of each professional group to discover or better describe eventual differences in timbre judgement based on music profession of respondents [7]. The embedding of individual dimensions of optimal models of professional groups into common perceptual space was successful (highly significant correlation coefficients), angles contained among individual embedded dimension coordinates showed only slight deviations from 90 o [12]. Comparing systems of (nearly) orthogonal verbal attributes which interpret common and group perceptual spaces implies following dimensional verbal attribute system for the description of musical sound timbre in Czech language: 1. temný / tmavý jasný / světlý ( / dark / bright in English) 2. drsný / hrubý jemný ( / rough ) 3. plný / široký úzký ( / wide ) Simplified common space scheme with representative attributes for each dimension selected according to their position in frequency vocabulary are in Figure 1. 3.3 Experiment 3 Twenty experienced musicians took part in the test of attribute suitability. Mean values of judgements were selected for the presentation of results which well represent the differences in attribute judgements. Mean values revealed that dimensional attributes, dark,,, rough,,, wide and found in experiment 2 and attribute are considered more then appropriate for the description of musical sound and (except rough) are more then often used (question number one). Answers for questions number three and four for individual attributes are summarized in Figure 2. 4 Discussion The results of the experiment 1 showed large variety of attributes in individual vocabularies (61 % of attributes were used only once) and also the stable and solid core of common vocabulary. This core contains the attributes with similar ranks in common and group vocabularies and also the dimensional attributes found in subsequent experiment. The application of CLASCAL procedure on dissimilarity results of all 43 respondents created three-dimensional common perceptual space of verbal attributes. The aposteriori classification of respondents is not connected with their music profession. The embedding of perceptual space dimensions of music professional groups into common perceptual space was highly successful and revealed only small deformation of angles among pairs of embedded dimensions [12]. Since some respondent group optimal solutions were only two-dimensional, we may conclude that at least these two dimensions are shared by five investigated professional groups: players of string, wind and key instruments, composers & conductors and sound designers. That means that there is no substantial difference among opinions of Czech speaking musicians on relations among verbal attributes describing musical sound timbre. Comparison of dimensions of common perceptual space with the dimensions postulated by Stumpf [4] (see also the Introduction) shows good agreement of the first and third dimension. Attribute, which is prominent attribute in almost all timbre studies with sound contexts, is missing in our second dimension. Its position in perceptual spaces of all respondents and professional groups is similar and lays closely to the plain defined by the first two dimensions in quadrant defined by attributes and (see Figure 1). Figure 1: Schematic diagram of common space of verbal attributes with dimensional attributes and. 1604
rather steady state a) Attack or steady state description (question No. 3) rather attack both vibratory buzzy grumpy brisk wide colored unstable honey cool warm sweet dark thudy bright hissy thin nasal damped flat carrier hollow bleaty blurred balanced rustle hearty radiant lucid silky sparse piping glossy velvety rumbly round voiced rude metallic pure lively sophisticated racy dull huskily ringing rough brazen unpointed pronounced piercing soft definite hard very appropriate b) Isolated tone timbre description (question No. 4) unappropriate adequate velvety ringing dark radiant blurred bright metallic lucid round pure bleaty colored hissy soft thin piping honey rustle buzzy voiced sweet rumbly nasal grumpy hollow rude pronounced glossy damped brazen silky warm wide cool hearty carrier piercing unpointed huskily hard rough thudy vibratory dull racy brisk balanced sparse flat definite sophisticated unstable lively Figure 2: Mean values of answers for (a) suitability of attribute for description of attack or steady state part of the tone, (b) suitability of attribute for description of timbre of isolated (standalone) tone of musical instrument. Values of dimensional attibutes and attribute are drawn with filled bars. very appropriate dimension 1 dimension 2 dimension 3 Suitable for description of: adequate timbre quality of playing unappropriate dark bright rough bbbbbbbbbbb wide artistic expression Figure 3: Mean values of answers for questions number four to six for attribute and dimensional attributes. 1605
Results of experiment 3 implies that attribute and dimensional attributes are appropriate and often used for the description of musical sound. This fact supports their importance among other studied attributes. The attributes representing the first and third dimensions are more suitable for the description of steady state part of the sound; attribute and the second dimension attributes are suitable for both attack and steady state (see Figure 2a). All dimensional attributes except wide and rough are very appropriate for description of timbre of isolated tone, attributes wide and rough are only adequate (Figure 2b). In all attributes representing the first dimension dominated suitability for the description of timbre of isolated tone (Figure 3) we may call the first dimension as timbre dimension. In the second dimension attributes dominated suitability for the description of quality of play of tone together with artistic expression we may call the second dimension as quality of play and artistic expression dimension. In attributes of the third dimension dominated suitability for the description of timbre and quality of playing of the tone we may call the third dimension as timbre and quality of play dimension. Attribute according to questions four to six behaved similarly to attributes of the first and third dimensions. 5 Conclusion Consistency, robustness, quality of results and also their similarity with previous results approved rightfulness and applicability of used approach in research of timbre to ask respondents for their opinions and experience about timbre without any sound context. Thus the found verbal attributes and their discovered relations can be used in future listening tests with various sound contexts. Future study of verbal attributes used for the description of timbre aims at two ways: in semantic or linguistic analysis of dimensional (most important) attributes found in Czech language and in search for analogous structures of attributes in other languages. 6 Acknowledgements The research was supported by the Grant Agency of the Czech Republic (Project No. 202/02/1370) and by the Ministry of Education and Youth of the Czech Republic (Project No. 1M6138498401). References [1] Grey, J. M.: Multidimensional perceptual scaling of musical timbres. Journal of the Acoustical Society of America 61 (1977), 1270-1277. [2] Bismarck, G. von: Timbre of steady sounds: A factorial investigation of its verbal attributes. Acustica 30 (1974), 146-159. [3] Stepanek, J.: The Study of Violin Timbre Using Spontaneous Verbal Description and Verbal Attribute Rating. In Forum Acusticum Sevilla 2002 (ISBN 84-87985-06-8): MUS-06-008. [4] Stumpf, C.: Tonpsychologie I-II. S. Hirzel Verlag, Leipzig 1883. [5] Moravec, O., Stepanek, J.: Verbal description of musical sound timbre in Czech language. In Proceedings of the Stockholm Music Acoustics Conference (SMAC'03), Stockholm, 643-645. [6] Moravec, O., Stepanek, J.: Collection of Verbal Descriptions of Musical Sound Timbre in Czech Language. In Proceedings of the 7 th International Colloquium "ACOUSTICS '03", Zvolen Šachtičky, 23-26. [7] Moravec, O., Stepanek, J.: Perceptual spaces of verbal attributes used for description of musical sound timbre in Czech language. In Proceedings of 7. CFA / 30. DAGA, Strasbourg, 881-882. [8] McAdams, S., Winsberg, S., Donnadieu, S., De Soete, G., Krimphoff, J.: Perceptual scaling of synthesized musical timbres: common dimensions, specificities, and latent subject classes. Psychological Research 58: 177-192. [9] Winsberg, S., De Soete, G.: A latent class approach to fitting the weighted Euclidean model, CLASCAL. Psychometrika 58: 315-330. [10] Borg, I., Groenen, P.: Modern Multidimensional Scaling, Theory and Applications. Springer- Verlag, New York. [11] Stepanek, J.: Relations between perceptual space and verbal description in violin timbre. In Acústica 2004 Guimarães, Portugal, CD ROM: AFP 077-S. [12] Stepanek, J., Moravec, O.: Verbal description of musical sound timbre in Czech language and its relation to musicians profession and performance quality. In Proceedings of the Conference on Interdisciplinary Musicology (CIM05), Montréal, 2005 1606