Tonal Hierarchies and Rare Intervals in Music Cognition

Long-Term Representations of Pitch Structure vs. Short-Term Effects of the Perceptually Immediate Context One issue raised by Butler (1989; see also Butler 1988b) is whether experiments such as these which demonstrate hierarchical differentiation of This content downloaded from 128.84.127.82 on Wed, 3 Apr 2013 15:52:11 PM All use subject to JSTOR Terms and Conditions 312 Carol L. Krumhansl tones reflect long-term knowledge of pitch structure or simply special strategies developed in response to the particular task and stimulus materials. He is concerned primarily with the probe-tone rating task, in which a context is followed on different trials by all 12 tones of the chromatic scale. To be concrete, Butler claims that the "typical" contextual sequence shown in his Figure 7 (which, to my knowledge, has never been employed in any published experiment) might induce a strategy to rate the tonic most highly because it is sounded twice, once at the beginning of the scalar sequence and once at the end, and with longer durations than the remaining diatonic tones. Even if these factors were influential, it would not account for the results he reproduces in his Figure 2. These results show differentiation among the remaining (nontonic) scale tones (e.g., higher ratings of the dominant than the subdominant), whereas his account would predict these to be equal. Arguments of this kind that probe-tone ratings simply reflect influences of such factors as tone repetition, duration, and serial position are unsupportable. The contexts actually used in the probe-tone studies (e.g., Krumhansl & Shepard, 1979; Krumhansl & Kessler, 1982; Cuddy & Badertscher, 1987) are varied: incomplete ascending and descending scales (without the final tonic), complete scales (with the final tonic), tonic triads (blocked and arpeggiated), and chord cadences. Despite these variations, the agreement between results from the different conditions evidences a robust and reliable pattern that cannot be attributed to the particulars of the contexts. In fact, the different contexts were used in the experimental designs for the purpose of testing for agreement in the rating data despite variations in the particular contexts used. Krumhansl and Kessler (1982), for example, reported a high correlation between the probe tone ratings for four different major key contexts (.896) and four different minor key contexts (.910). These values were considerably higher than would be predicted from correlating the tone distributions in the different contexts (the correlations average .752). This shows that the consistency of the data cannot be attributed to stimulus factors alone. Had Butler done this appropriate analysis, rather than summing tone distributions (his Figure 8), he would have been unable to conclude "this similarity [between probe tone ratings and tone distributions] reflects a stimulus artifact in the design of Krumhansl and Kessler's experiment, rather than a reflection of mental representations of pitch relations" (pp. 230-231). Butler also disregards data arising from experiments that use different tasks. For example, tonal hierarchies are evident also in memory confusions (e.g., Cuddy, Cohen, & Miller, 1979; Krumhansl, 1979), reaction times to judge key membership (Janata & Reisberg, 1988), judgments of what constitutes a good or complete phrase (Palmer & Krumhansl, 1987a,b), and expectations for melodic continuations (Schmuckler, 1988). The last three This content downloaded from 128.84.127.82 on Wed, 3 Apr 2013 15:52:11 PM All use subject to JSTOR Terms and Conditions Tonal Hierarchies and Rare Intervals 313 studies, as well as the probe-tone experiment of Thompson (1986), employ segments of musical compositions, demonstrating that these evoke similar abstract tonal hierarchies. In short, Butler's conclusion that "serious problems" of a methodological nature attend the "atonal hierarchy theory" are readily disconfirmed by considering the extent to which short-term effects of the perceptually immediate context fail to predict the experimental data at hand, and the range of tasks and contexts generating a similar pattern of hierarchical differentiation of tones. Butler's (1989, p. 233; see also Butler, 1988b) account of the second probe-tone study of Krumhansl and Kessler (1982) reveals a number of misconceptions about how the data in that study were analyzed and interpreted. These aspects are important for understanding the approach taken to evaluate the extent to which the results reflect local effects of individual chords. More generally, they illustrate how the probe-tone methodology can be used as a tool for uncovering different levels of musical organization. The objective of the study was to trace how the sense of key develops and changes over time as sounded information is evaluated and integrated by the listener. For this purpose, we obtained probe-tone ratings after each successive chord in various sequences, some of which suggested modulations to more or less closely related keys. The ratings at each position were correlated with the tonal hierarchies obtained separately for unambiguous key-defining contexts; these values (rather than "rankings of well-fittedness of probe tones") were taken as the measure of key strengths. These were compared with analogous correlations by using probe-tone ratings for isolated chords also obtained separately. (This is how we used the data Butler, 1989, p. 229, footnote 10, claims "were dropped from the analysis without explanation.") These two sets of values are plotted in Figure 8 of Krumhansl and Kessler (1982), where it is explained: "The difference between the graphs ... is a measure of the extent to which the listener perceives the prevailing tonal organization as opposed to the tonal organization of the last heard chord" (p. 356). Treating the data in this way enabled us to separate effects of tonicization at one level from a sense of the prevailing key at another level. This was related to Schenker's (1906/1954, 1935/1979) theory of hierarchic levels of tonal organization. Thus, it seems unfair to conclude that the probe-tone method is unable to separate surface from more abstract levels of organization (Butler, 1989, p. 223) or to account for the process of tonicization (Butler, 1989, pp. 239-240). This and other studies show that the perceptually immediate context has an effect on the musical experience. It would be strange, indeed, if the sounded events did not have special salience. What the experimental studies reviewed to this point demonstrate is that, in addition, tonal contexts (even simple, schematic ones) evoke listeners' knowledge about how musical enThis content downloaded from 128.84.127.82 on Wed, 3 Apr 2013 15:52:11 PM All use subject to JSTOR Terms and Conditions 314 Carol L. Krumhansl tities, such as tones, chords, and keys, are generally used in the style. These abstract pitch relations provide a background that facilitates understanding the organization of a particular piece: the roles of sounded events in larger melodic, harmonic, and rhythmic patterns; the building and release of tension; modulations to more or less closely related keys; and so on. As described by Bharucha (1984b), each event has a place not only in the abstract tonal and harmonic hierarchies (referring to classes of tone and chord instances), but also a place in the event hierarchy of the particular piece (referring to the string of musical events that constitutes a piece of music). Moreover, a number of studies suggest that surface salience (resulting from such factors as tone repetition, duration, overall pitch height, serial position, metrical stress) may be important for enabling the listener to discover principles of pitch organization in novel musical styles. For example, Castellano, Bharucha, and Krumhansl (1984) found style-appropriate tonal hierarchies for North Indian music with inexperienced listeners, which was attributed to the explicit emphasis in the music given to structurally significant tones (as summarized in tone durations). Similar findings were obtained in Kessler, Hansen, and Shepard's (1984) study of Balinese music. Effects of surface salience also were found in studies of octatonic (Krumhansl & Schmuckler, 1986b) and 12-tone serial writing (Krumhansl, Sandell, & Sargeant, 1987). In all these cases (with the last being a possible exception), surface salience appears to initiate the abstraction of styleappropriate pitch organizations in music outside the listener's prior experience. With increased exposure, a process of this sort is assumed to establish long-term representations of compositional principles that can be evoked with simple, schematic contexts. Statistical Distribution of Musical Elements and Psychological Representations of Pitch Structure The next issue to be discussed is the utility of statistical surveys of musical compositions in connection with psychological research. The correspondence between the profile of total durations of the 12 chromatic scale tones in Schubert, op. 91 no. 1 (Hughes, 1977) and psychological measures of the tonal hierarchy is striking; the correlation between the tone durations and the tonal hierarchy of G major is .97. Hughes (1977, p. 157) states: "Considered as a whole, the results of op. 94 no. 1 indicate a rather simple tonal complexity much closer to the key of G major than to C major as presented by Schubert." Butler's (1989, p. 225) confusion on this point ("Hughes (1977) stated that his sums of note durations indicated that the music was oriented "toward" G major although the key of the composition was C major, and indeed Schubert seemed to believe the piece was in C maThis content downloaded from 128.84.127.82 on Wed, 3 Apr 2013 15:52:11 PM All use subject to JSTOR Terms and Conditions Tonal Hierarchies and Rare Intervals 3 15 jor, as well.") may stem from Butler's disregard for Hughes's distinction between tonal organization (based on a chord-by-chord analysis) and tonal orientation (based on the total occurrence of each note and its durational value, which are treated with a number of analytic tools). Whether or not this distinction is theoretically useful, the long central section of the piece has the key signature of G major. Hughes also analyzes shorter sections of the piece separately, finding different tonal regions. In a corresponding way, the tonal hierarchies of different keys provide the best match to the tonal distributions in the separate sections. The analysis of tone distributions in this particular piece raised the question of the generality of the correspondence between distributions of elements in music and psychological data on perceived tonal and harmonic structures. As summarized in Krumhansl (1987b; and more extensively in Krumhansl, 1990), the correspondences are generally strong. Figure 1 shows the tone frequencies in a variety of melodies (a sample of nearly 25,000 tones) from studies by Youngblood (1958) and Knopoff and Hutchinson (1983) and the tonal hierarchies (Krumhansl & Kessler, 1982) of the appropriate key. (In the figure, the values are plotted with respect to a reference tonic of C.) Similar correspondences obtain between frequencies of successive pairs of tones and judgments of perceived relatedness between tones, frequencies of chords and perceived harmonic hierarchies, and frequencies of harmonic progressions and judgments of perceived relatedness between chords. These analyses show a strong, general correspondence between distributions of elements in musical compositions and psychological data. Naturally, the relationships would be weaker if, for example, the statistical sample were based on a composition involving complex modulations, on short segments for which the distributions would be less stable, or on a segment of a composition intended to set up conflicting or ambiguous tonal cues. What significance do these correspondences have for music perception? Butler (1989, p. 222) states: "The tonal hierarchy theory has been invoked to account for ... statistical distributions of tones within actual musical compositions (Krumhansl, 1986 [Krumhansl, 1987b])." In fact, just the opposite was proposed: "These correspondences suggest that knowledge of tonal practice evident in listeners' judgments is abstracted from their musical experience, and that through this experience they have internalized the regularities of the musical style" (Krumhansl, 1987b, p. 22). This hypothesis is congruent with a wide variety of results showing that humans (as well as other organisms) are highly sensitive to information about frequency of occurrence. Thus, the primary significance of the observed correspondence between statistics of music and psychological data in these cases is to suggest a mechanism through which principles of musical organization are learned. This content downloaded from 128.84.127.82 on Wed, 3 Apr 2013 15:52:11 PM All use subject to JSTOR Terms and Conditions 316 Carol L. Krumhansl Fig. 1. The distribution of tones (the total frequency of occurrence) in tonal compositions (from Youngblood, 1958; Knopoff & Hutchinson, 1983) compared with the major key profile (top) and the minor key profile (bottom). Correlations between tone distributions and perceived tonal hierarchies are strong for both major and minor keys. (Reprinted with permission from Krumhansl, 1990.) Statistical summaries also can guide psychological experimentation by helping to identify typical (as opposed to infrequent) pitch patterns for study. For example, a glance at Figure 1 (as well as the distributions presented by Butler, 1989) shows that the diatonic tritone in major keys (between the subdominant and the leading tone) is much less frequently represented in musical compositions than most all other diatonic intervals. This raises the question of what process might enable listeners to learn the speThis content downloaded from 128.84.127.82 on Wed, 3 Apr 2013 15:52:11 PM All use subject to JSTOR Terms and Conditions Tonal Hierarchies and Rare Intervals 317 cial significance of the tritone as a cue for the key of a piece of music. Finally, statistical summaries can help relate musical and psychological constructs to one another. The literature contains a number of successful applications of this approach. Lindblom and Sundberg (1970) used the rate of tonal events to substantiate a generative theory of constituent structure. Hutchinson and Knopoff (1979) related the predominance of chord voicings with larger intervals in the bass than the treble to increases in critical band width in lower registers. Recently, Vos and Troost (1989) found ascending intervals tended to be larger than descending intervals, consistent with Meyer's (1973) gap-fill melodic process and the idea that the large ascending interval induces a feeling of tension or arousal that is reduced when the gap is subsequently filled in by smaller descending intervals. Although surveys of features found in musical compositions can only provide information about the way elements are generally used in a musical style, they can serve to substantiate the operation of certain psychological processes. The Process of Key-Finding A variety of psychological measures indicate that listeners experienced with tonal-harmonic music tend to interpret the sounded tones and chords in terms of their functions in the prevailing key. This raises the question of how listeners initially arrive at a sense of key. Intuitively, it seems that certain cues may be important: chord cadences, tonic-dominant-mediant relations, the diatonic set, and the "rare" intervals of tritones and minor seconds. In evaluating these possibilities, it seems important to bear in mind that they are not mutually exclusive. Listeners may use a number of cues in combination, with the weighting possibly dependent on the musical passage in question. The correspondence described in the last section between tonal distributions and tonal hierarchies suggests that listeners may develop a sense of key, in part, by matching the sounded events in a particular passage to the learned tonal hierarchies. A computer algorithm (Krumhansl & Schmucker, 1986a; Krumhansl, 1987b, 1990) was developed to test the efficacy of a key-finding process of this sort. This simple algorithm is based on matching the distribution of tone durations in the input segment to the tonal hierarchies of the 24 major and minor keys. This is done by correlation, which matches the patterns as wholes rather than by piecemeal comparison of relative rankings as was done by Butler (1989, pp. 226-228). The key with the highest correlation is said to be the key found by the algorithm. Essentially, this algorithm simulates a process of patternmatching between the durations of the tones in a musical excerpt and the psychological tonal hierarchies. In tests of the algorithm, the tone durations were taken from the notated score as a simple measure of the tonal content This content downloaded from 128.84.127.82 on Wed, 3 Apr 2013 15:52:11 PM All use subject to JSTOR Terms and Conditions 318 Carol L. Krumhansl of the passage. This measure ignores all other factors such as temporal order, harmonic implications, and metrical and rhythmic stress. How these factors might be incorporated in such an algorithm to sharpen the keyfinding process is discussed in Krumhansl (1990). Various theoretical constructs, such as prominence profiles (Lindblom & Sundberg, 1970) and reductions (Lerdahl & Jackendoff, 1983), also may prove useful in filling out a description of the key-finding process. Even without taking these other factors into account, tests of the algorithm showed that matching to tonal hierarchies yields fairly accurate results and this is true for very short musical excerpts. For purposes of later comparisons with a tritone rule (applicable only to major keys), I will focus on the applications of the algorithm to pieces in major keys. One test of the algorithm used just the first four sounded tones in the preludes from Bach's Well-tempered Clavier, and the preludes of Shostakovich and Chopin. The algorithm determined the correct key for 79% of the major preludes using just the first four tones. In an application to the fugues of Bach and Shostakovich, the tones of the fugue subjects were added one-byone. The correct major key was determined in 92% of the cases, with an average of 3.7 tones needed. A final application showed the algorithm successfully traced the shifting pattern of tonalities in Bach's C minor prelude (Book II) in a manner that agreed well with judgments of key strengths by trained music theorists; in this case, the analysis was done on a measure-bymeasure basis. Thus, tonal distributions in short excerpts contain a great deal of information about the key when matched to tonal hierarchies, contrary to Butler's (1989, p. 226) claim that, based on statistical distributions of tones and tonal hierarchies, "a listener could identify the key of a piece of music with full certainty only after having heard and summed the durational weightings of all the tones in the composition that is, after the performance was over." How effective might a tritone rule be for determining key? Butler and Brown (1984) reported an experiment in which listeners were 83-91% accurate in identifying the tonic after hearing three-tone sequences including the unique tritone of a major key (formed by the subdominant and leading tone). They conclude (p. 12) the result "provides persuasive evidence that the tritone serves as a primary indicator of tonal center," ignoring the finding that the same listeners were 96-98% accurate in choosing a possible tonic when the tritone was absent. When Butler (1989, p. 234) says of this study that "agreement erodes when rare intervals are absent," it is presumably simply because the three-tone sequences without tritones allow multiple key interpretations. To measure correctly the contribution of the tritone, what is needed is a design in which two possible alternative tones are presented, one a possible tonic and one an impossible tonic. If the percentage of correct tonic judgments is higher for tritone sequences, this would support Butler and Brown's (1984) conclusion. Unfortunately, a design This content downloaded from 128.84.127.82 on Wed, 3 Apr 2013 15:52:11 PM All use subject to JSTOR Terms and Conditions Tonal Hierarchies and Rare Intervals 319 such as this has not been employed. Other studies (e.g., Butler, 1982, 1988a) do not permit the contribution of the tritone to be evaluated, because suitable control conditions without tritones, essential for testing the theoretical claim, are lacking. (Incidentally, when Butler, 1989, p. 239, reports of the 1988a study "responses to patterns derived from tonal compositions approached unanimity for every pattern but one," he is talking about a total of three patterns. In addition, the motivation for including three atonal selections in a study of key-finding is obscure.) Let us suppose for sake of argument, however, that a tritone rule might be one mechanism listeners use to arrive at a sense of key. It has the appeal of providing potentially definitive information (assuming major mode and no chromatic alterations). One can then go on to ask how often the music would offer an opportunity to apply such a rule. In the initial four-tone segments of the Bach, Shostakovich, and Chopin major preludes the tritone appears just once that is, in 2% of the cases. [That Cohen's (1977) listeners were 83% accurate judging the key of a subset of the Bach major preludes after hearing just the first four events eliminates the tritone as necessary for key-finding.] In only 58% of the major fugues of Bach and Shostakovich do the two tones of the tritone appear anywhere (not necessarily successively) in the entire fugue subject. When they do appear it is after an average of 12.2 tones. Thus, opportunities to apply a tritone rule occur relatively late and infrequently in the initial segments of these pieces, even though "Everyday musical experience tells us that tonally enculturated listeners can recognize the tonal center in an unfamiliar composition almost instantly" (Butler, 1989, pp. 237-238). Other observations suggest the applicability of a tritone rule is limited: 1. It can rarely be applied to monophonie music [according to Butler, (1988a) tritones appear predominantly as harmonic intervals; the tritone constitutes only 0.3% of Youngblood's (1958) sample of nearly 2,000 melodic intervals]. 2. It cannot be applied unambiguously to pieces in minor keys (the minor scale in both harmonic and ascending melodic forms contains two tritones). 3. It cannot be applied to other church modes or scales in other cultures (e.g., 4 of the 10 thäts that is, scales on the circle of thäts in North Indian music do not have a unique tritone; see Krumhansl (1987a)]. 4. The tritone is relatively difficult to encode, label, remember, and produce. 5. A process keying-off (as it were) from tritones would be subject to false alarms when chromatic alterations occur [the only tonal example offered by Butler (1989), Figure 16, is a case in