Levitin, D. J. & Cook, Perry R.(1996) Memory for musical tempo: Additional evidence that auditory memory is absolute. Perception & Psychophysics, 58, pp. 927-935.


Memory for musical tempo: 

Additional evidence that auditory memory is absolute

Daniel J. Levitin

University of Oregon, Eugene, Oregon

and Stanford University, Stanford, California

and

Perry R. Cook

Stanford University, Stanford, California


This is an electronic Web version of the paper originally appearing in Perception & Psychophysics, 1996, 58, 927-935. Copyright 1997 Daniel J. Levitin. Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted with or without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and full citation on the first page. To copy otherwise,to republish, to post on services, or to redistribute to lists, requires specific permission and/or a fee.


Abstract

We report evidence that long term memory retains  absolute (accurate) features of perceptual events. Specifically, we  show that memory for music seems to preserve the absolute tempo  of the musical performance. In Experiment 1, 46 subjects sang  popular songs from memory, and their tempos were compared to  recorded versions of the songs. Seventy-two of the subjects came within  8% of the actual tempo on two consecutive trials (using different  songs), demonstrating accuracy near the perceptual threshold (JND)  for tempo. In Experiment 2, a control experiment, we found that  folk songs lacking a tempo standard generally have a large variability in tempo; this counters arguments that memory for the  tempo of remembered songs is driven by articulatory constraints.  The relevance of the current findings to theories of perceptual  memory and memory for music are discussed.   


A fundamental problem facing memory theorists is how to  account for two seemingly disparate properties of memory. On  the one hand, a rich body of literature suggests that the role of  memory is to preserve the gist of experiences; memory functions  to formulate general rules and create abstract concepts on the  basis of specific exemplars (Posner & Keele, 1970; Rosch, 1975).  On the other hand is the extensive literature suggesting that  memory accurately preserves absolute features of experiences  (Brooks, 1978; Jacoby, 1983; Medin & Schaffer, 1978). (For a  further discussion of these two perspectives, see McClelland and  Rumelhart, 1986).  

These perspectives on the function of human memory  parallel an old debate in the animal-learning literature about  whether animals' internal representations are relational or  absolute (Hanson, 1959; Reese, 1968). As with many  psychological debates, there may be some degree of truth on both  sides. Currently, philosophers of mind (Dennett, 1991) and  cognitive scientists (Kosslyn, 1980) have also wondered to what  extent our mental representations and memories of the world are  perfect or accurate copies of experience, and to what extent  distortions (or generalizations) intrude. 

The study of memory for music is potentially helpful in  exploring these issues because the experimental evidence is that  both relational and absolute features of music are encoded in  memory. Researchers have shown that people have little trouble  recognizing melodies transposed in pitch (Attneave & Olson,  1971; Dowling & Bartlett, 1981), so it is clear that memory for  musical melody must encode the abstract information, that is,  the relation of successive pitches, or pattern of tones if not  their actual location in pitch space.  

Abstract encoding has also been demonstrated for temporal  features: people easily recognize songs in which the relation  between rhythmic elements (the rhythmic pattern) is held  constant, but the overall timing or musical tempo has been  changed (Monahan, 1993; Serafine, 1979). (We define tempo as  the "pace" of a musical piece, or the amount of time it takes a  given note or the average number of beats occurring in a given  interval of time, usually beats per minute.)  

With respect to this point about musical relations, Hulse  and Page (1988) explain,  

    music emphasizes the constancy of relations among  sounds...within temporal structure, music emphasizes the constancy of relations among tone  durations and intertone intervals. Rhythmic structures  remain perceptually equivalent over a broad range of  tempos...and tempo changes involve ratio changes of  duration and interval. (p. 431).  

And, as Monahan (1993) argues, musical pitch and time are described most naturally in relative rather than absolute terms.  Thus (within very broad limits), the identity and recognizability of a  song is maintained through transposition of pitch and changes in  tempo. 

Evidence that memory for music also retains absolute pitch information alongside an abstract melody representation has  been mounting for some time (Deutsch, 1991; Halpern, 1989; Levitin,  1994; Lockhead & Byrd, 1981; Terhardt & Seewann, 1983). For example, in a previous study one of us (Levitin, 1994) asked  experimental subjects (most of them non-musicians) to sing their favorite rock and roll songs from memory, without any external  reference. When compared to the actual keys of the songs, the  subjects' productions were found to be at or very near the songs'  actual pitches. This is somewhat surprising from the standpoint of  music theory (and perhaps from "object perception" theories)  because that which gives a melody its identity or "objectness" is the  relation of successive intervals (both rhythmic and melodic). As  most music theorists would agree, "music places little emphasis on  the absolute properties of sounds, such as their exact temporal pitch  or duration" (Hulse & Page, 1988, p. 432). 

The animal learning literature reveals parallel evidence for  both abstract and relational memory for auditory stimuli. Starlings  (Hulse & Page, 1988) and whitethroated sparrows (Hurly, Ratcliffe,  & Weisman, 1990) were shown to use both absolute and relative  pitch information in discriminating melodies, but other evidence is  mixed. Barn owls seem to remember musical patterns based on  absolute pitch (Konishi & Kenuk, 1975), and wolves use absolute  pitch information to the exclusion of relative pitch information to differentiate between familiar and unfamiliar howls (Tooze,  Harrington, & Fentress, 1990).  

What reasons are there to expect that long term memory might  encode tempo information with a high degree of accuracy? Scores of conditioning experiments with animals have shown that animals can  learn to estimate interval durations with great accuracy (e.g., Hulse,  Fowler, & Honig, 1978).  

The question of human memory for tempo has been addressed in  several key studies. People report that their auditory images seem  to have a specific tempo associated with them (Halpern, 1992), and  in one study, subjects who imagined songs tended to imagine them at  about the same tempo on occasions separated by as much as five  days (Halpern, 1988). Collier & Collier (1994) found that trained  jazz musicians tended to vary tempo less than 5% within a song or  across multiple performances of the same song on different days,  suggesting a stable memory for tempo in these musicians. 

Whereas the studies just mentioned argue for the stability of  internal tempos, the question of the objective accuracy of these internal tempos interested us. That is, when we remember a song  (or "auralize" it, to use Ward's terminology) do we do so at its original tempo? Just as some people can match pitches accurately  (and we say they have "absolute pitch"), we wondered if there are  those who can match tempos accurately, and have an ability of  "absolute tempo." In particular, we wanted to study this ability in  everyday people with little formal musical training. This question  is relevant to researchers studying human rhythm perception (e.g.  Desain, 1992; Jones & Boltz, 1989; Povel & Essens, 1985; Steedman,  1977), the nature and stability of internal clocks (Collier & Wright,  1995; Collyer, Broadbent, & Church, 1994; Helmuth & Ivry, 1995),  mental imagery (Finke, 1985; Kosslyn, 1980) and general theories of  time perception (Block, 1990; Fraisse, 1981; Michon & Jackson, 1985).  

On a music-theoretic level, Narmour (1977) argues that  musical listening involves using both schematic reduction and  unreducible (absolute) idiostructural information. To the extent  that the original listening experience is preserved in the brain, we  might expect to find these two types of information represented in  memory, not just in perception. One of the goals of the current  study was to test this hypothesis. 

EXPERIMENT 1 

Experiment 1 was designed to discover if people encode the  absolute tempo of a familiar song in memory, and if so, with what  degree of precision. In Levitin (1994), subjects were asked subjects  to sing contemporary popular and rock songs from memory and their  productions were analyzed for pitch accuracy. Using these same  data, we analyzed their previous productions for tempo accuracy.  Contemporary popular and rock songs form an ideal stimulus set for  our study because they are typically encountered in only one version  by a particular musical artist or group, and so the song is always  heard - perhaps hundreds of times - in the same key, and at the same  tempo. (In fact, we selected out songs that did not meet this criterion.)  

METHOD.  The raw data used in this study were originally collected for a  study on pitch memory (Levitin, 1994). 

SUBJECTS.  The subjects were 46 Stanford University students who served  without pay. The subjects did not know in advance they were  participating in a study involving music, and the sample included  subjects with and without some musical background. All subjects  filled out a general questionnaire before the experimental session.  The subjects ranged in age from 16 to 35 years (mean, 19.5; mode,  18; SD, 3.7).  

By self-report, the subjects' musical background ranged from  no instruction to more than 10 years of instruction; 37 subjects had  some exposure to a musical instrument, 9 had none. In response to  the question "how much structured musical training in either  performance or theory have you had?" 17 subjects reported none; 17  subjects reported 1-3 years; 5 subjects reported 3-5 years; 3  subjects reported 5-7 years; 3 subjects reported 7-10 years; and 1  subject reported more than 10 years. 

MATERIALS.  Prior to data collection, a norming study was conducted to  select stimuli with which this subject population would be  familiar. 250 introductory psychology students completed a  questionnaire which asked them to indicate songs  that "they knew well and could hear playing in their heads." None of  the subjects in the norming study were subsequently used in the main  experiment. 

The results of this norming study were used to select the best  known songs. Songs on this list that had been performed by more than  one group were excluded from the stimulus set because of the  possibility that these versions might have different tempos. From this questionnaire, fifty-eight compact discs (CDs) were selected, representing over 600 songs  from which the subjects could choose. Examples of songs included  are "Hotel California" by The Eagles; "Get into the Groove" by  Madonna; and "This & That" by Michael Penn. (A complete list of CDs  constituting the stimulus set is available from the first author.) 

PROCEDURE.  Subjects were seated in a sound attenuation booth alongside  the experimenter. The 58 CDs chosen from the norming study were  displayed alphabetically on a shelf in front of the subjects. The  experimenter followed a written protocol asking subjects to select  from the shelf and to hold in their hands a CD that contained a song  they knew very well. Holding the CD and looking at it may have  provided a visual cue for subsequent auditory imaging. There was no  CD player in the booth, and at no time were the CDs actually played  for the subjects. All subjects reported that they had not actually  heard their chosen song in the previous 72 hours, and many had not  heard it in months. 

The subjects were then asked to close their eyes and imagine  that the song was actually playing. They were told that, when they  were ready, we wanted them to try to reproduce the tones of the  song by singing, humming or whistling, and they could start  anywhere in the song they wanted to. The subjects were not  explicitly told anything about rhythm or tempo, nor were they  specifically asked to reproduce the tempo of the songs. Their  productions were recorded on digital audio tape (DAT) so that pitch  and speed would be accurately preserved. The subjects were not  told how much of the song to sing, but they typically sang a four-bar  phrase. Following the production of this first song, the subjects  were asked to choose another song and repeat the procedure; this  constituted the two experimental Trials. Three of the subjects  discontinued their participation after Trial 1.  

ANALYSIS.  The subjects' productions were compared with the songs  performed by the original artists on CD, in order to compare tempos.  The subjects' productions and the corresponding sections of the CD  were transferred digitally to a MacIntosh computer, and the sample  rate was converted from its original 44.1 KHz or 48 KHz to 22.050  Khz for storage economy.  

The duration of each subject's production and the associated  CD excerpt were measured using the program MacMix, and these  measurements were accurate to within 0.1 msec. A typical subject  production was 5 seconds long. The total selection was also divided  into beats yielding two equivalent measures of production time:  total duration and beats per minute. For the purposes of this report,  data are presented as tempos in units of beats per minute. 

RESULTS.  Figure 1 shows, as a bivariate scatterplot, the tempos  produced by subjects compared to the actual tempos of the  remembered pieces (Trials 1 and 2 are combined). The subjects  came very close to their target tempos as indicated by the high  correlation between subjects' tempos and actual tempos (r=.95), and the fact that most responses fall near the  diagonal. Figure 2 shows the distribution of errors that the subjects  made, expressed in a histogram as percent deviations from the  actual tempo.1  


Figure 1. Subjects' tempos versus Actual tempos, both trials combined.


On Trial 1, 33/46, or 72%, of the subjects performed within +/- 4% of the  actual tempo for the songs, and 41/46, or 89%, of the subjects performed  within +/- 8% of the actual tempo (M = 4.1%; SD = 7.7%). On Trial 2,  12/42, or 40%, of the subjects came within +/- 4%, and 25/42, or 60%, came within +/- 8% of  the actual tempo (M = 7.7%; SD = 7.9%). As Figure 2 shows, for the  two trials combined, 72% of the responses fell within 8%. 

To put these results in context, one might ask what is the JND  for tempo? Drake and Botte (1993, Experiment 3) found the JND for  tempo discrimination to be 6.2% - 8.8% using a two-alternative  forced choice listening ("which is faster?") test; Friberg & Sundberg  (1993) found JND for tempo to be 4.5% using the psychophysical  method of adjustment; Hibi (1983) found the JND to be ~6% for  displacement of a single time marker in a sequence, and for  lengthening/shortening of a single time marker. In tapping tasks,  where subjects had to either tap along with a pulse at a certain  tempo (a "synchronization task") or continue tapping to a tempo set  up by the experimenter ("continuation task"), JNDs of 3-4% have been  reported for synchronization (Collyer, Broadbent, & Church, 1994;  Povel, 1981), and 7-11% for continuation (Allen, 1975). All of the  above JNDs apply for tempos in the range our subjects sang. It  appears, then, that a large percentage of the subjects in our study  performed within one or two JNDs for tempo, based solely on their  memory for the musical pieces. 

A somewhat more ecologically valid confirmation of these JND  figures comes from Perron (1994). In contemporary popular music,  many recordings are made with drum machines or computer  sequencers instead of live drummers. The anecdotal opinion of  musicians and record producers has been that these machines are  much more able to hold a steady tempo than human players. Perron  measured the tempo deviations in a number of these devices and  found the mean tempo deviation to be 3.5% (with a standard  deviation of 4.5%). Because the deviations in these machines seem  to go largely unnoticed by most people (including professional  drummers), it seems fair to assume that this 3.5% is less than the  JND for tempo variation. An interesting implication of Perron's  finding is that our subjects may well have been trying to reproduce  tempos for songs that contain variations of this magnitude. 

One might ask whether the subjects in our study performed  consistently across the two trials. To measure this, trials on which  the subjects came within +/- 6% were considered "hits" and all others  were considered "misses," in accordance with the more conservative  of Drake & Botte's (1993) JND estimates. 25 subjects (or 60%) were  found to be consistent in their performance across trials. Yule's Q  was computed as a measure of strength of association for this 2x2  table, and was significant (Q = .50; p<.04). 

Next, we wondered whether the subjects who had accurate  tempo memory also had accurate pitch memory (as measured in  Levitin, 1994). Combining Trials 1 and 2, and using 6% as a "hit" criterion for tempo and +/- 1 semitones (s.t.) as a criterion for pitch,  there was not a significant association. (Yule's Q = .31; n.s.).2

One of the implicit assumptions in these analyses is that the  tempos of the songs our subjects sang are widely distributed. If all  the songs fell into a narrow tempo band, one might argue that the  subjects only have memory for a particular tempo (or narrow set of  tempos). As Figure 1 shows, however, the range of tempos produced  by subjects was very large, running from approximately 60 bpm to  over 160 bpm.  

Similarly, one might wonder if the good performance of  subjects across trials was merely due to all of the subjects singing  songs at the same tempo in both cases; individual subjects may have  an idiosyncratic "preferred" tempo (or "internal tempo") that they  know well and relied on for this task (Braun, 1927). We found the  correlation between tempos sung on Trial 1 and Trial 2 was very low  (r=.07), as was the correlation between "targets" on Trial 1 and Trial  2 (the tempos subjects were trying to reproduce; r=.04).  


Figure 2. Percent deviation from actual tempo, both trials combined.


Variations of +/- 5% during the subjects' learning of the material  could have occurred if subjects had heard the songs repeatedly on a  cassette player or record player that did not keep accurate speed; CD  players are not subject to speed fluctuations. A questionnaire item  asked the subjects about whether they had heard their chosen song  on CD, radio, cassette, or record player, and there was no correlation  found between the source of the learning and their performance (all  the commercial radio stations in the area of the study were  broadcasting CDs exclusively during the study period, so "radio"  responses were considered to be "CD learning"). Similarly no  correlation was found between other factors such as sex, age,  handedness, or musical training. 

DISCUSSION.  The finding that 89% of the Trial 1 subjects and 60% of the  Trial 2 subjects made errors of only +/- 8% is evidence that long term  memory for tempo is very accurate, and is near the discrimination  threshold (as measured by JNDs) for variability in tempo. These  results may even underestimate the strength of tempo memory,  because our subjects were only instructed to reproduce pitches  accurately; to the extent that they also reproduced tempo, they did  this on their own, and without being requested by the experimenter  to do so.  

The distribution of errors is also instructive. As Figures 1 and  2 show, there is a tight clustering near the  actual tempo, and more subjects sang too fast than too slow. Boltz  (1994) reviews evidence that various forms of induced stress  increase the internal tempo of individuals. This would create  internal durations that are shorter than the standard, and cause the  subjects to sing fast. If we can assume that the experimental  situation was somewhat stressful (many subjects seemed to be  embarrassed or nervous), this could account for the asymmetric error  distribution favoring faster reproductions. An additional explanation  for this asymmetry comes from experimental findings that people are  more likely to perform faster rather than slower (Kuhn, 1977), and  are better able to detect tempo decreases than increases (Kuhn,  1974). 

In spite of the certain awkwardness of being asked to sing for a  psychological experiment, and the concomitant desire to be done with  the task as quickly as possible, most subjects reproduced the tempo of their selected songs with  remarkable accuracy. In listening to the subjects' productions, and  the corresponding artists' renditions, we were struck by how close  the subjects came not just in tempo, but in pitch, phrasing, and  stylistic nuances while singing from memory. In many cases it  seemed that the subjects couldn't have performed better if they  were actually singing along with the CD - but of course, they were  merely singing along with a representation of the CD in their heads. 

One could argue that our findings are the result of an artifact  rather than actual "memory for tempo." While recalling these  musical pieces, the subjects sing or imagine lyrics, and perhaps the  lyrics provide a constraint for the tempo. That is to say, the tempo  of a piece might be constrained by the number of syllables that have  to be fit into a particular melody. To counter this argument, one  would need to find a piece of music with lyrics that has no well- defined tempo standard, and ask subjects to sing it. If the range of  tempos produced for such a song is much wider than the range  produced by our subjects, we could argue that articulatory  constraints do not account for memory for tempo.  

As a first look at this issue, we examined Halpern's (1988)  data. In her Experiment 1, Halpern asked subjects to imagine popular  songs and set a metronome to match the tempo they heard in their  heads. Most of the songs had the property that no single reference  (or canonical) version existed - for example, "Happy Birthday,"  "London Bridge is Falling Down," and "Twinkle, Twinkle, Little Star."  In general, peoples' exposure to these songs is through informal live  singing (such as in elementary school), and the variety of recorded  versions of these songs virtually ensures that there is no uniformly  "correct" tempo. Halpern provided us with the (previously  unpublished) standard deviations across subjects for these three  songs, and they were (respectively) 16%, 19%, and 22% (see Table 1).  

In a replication (Experiment 2), Halpern asked a second  group of subjects to perform the same task. The mean tempo (across subjects) for each song varied by a large amount from one  experiment to the other: for the three songs (respectively) the difference in mean tempos was 19%, 12%, and 14%. Halpern also  asked her Experiment 2 subjects to adjust a metronome  incrementally to the point that represented the fastest and slowest  they could imagine the song. As Table 1 indicates, the tempos  selected by the subjects varied more than 250%. All of Halpern's  tempo variations are larger than those we found in our subjects,  supporting our claim that the tempo was not tightly constrained by  the lyrics.  

As a control condition, we designed Experiment 2 to  replicate Halpern's (1988) earlier findings about tempo variability in a  production context. We recruited eight subjects and asked them to sing three familiar folk songs (mentioning nothing to them about  tempo), and afterwards to sing them as fast and slow as possible. A  large standard deviation in this task would replicate Halpern's  finding and indicate that lyrics do not significantly constrain tempo.

 EXPERIMENT 2

 METHOD. 

SUBJECTS.  The subjects were 4 University of Oregon students, and 4  members of the community, recruited without regard to musical  training; six subjects had no previous musical instruction, two  subjects had less than two years of musical instruction. All the  subjects served without pay. 

MATERIALS.  To investigate tempo variability in singing, the subjects  were asked to sing "Happy Birthday," "We Wish You A Merry  Christmas," and "Row, Row, Row Your Boat."  

PROCEDURE. The subjects were asked to sing one of the three songs (song  order was randomized). When they finished, they were next asked to  sing it "as slow as you possibly can" and then "as fast as you possibly  can." This was repeated for the other two songs. The subjects were  recorded either directly to the hard disk of a NeXT computer using the  program SoundEditor, or to a Sony DATMAN Digital Audio Tape  Recorder which was then transferred digitally to NeXT computer  sound files.  




RESULTS. Table 2 shows the distribution of tempos for the three songs  sung at their normal speeds. The standard deviations are all well  above the ~8% standard deviation we found for our Experiment 1  subjects. An F test between the variance in Experiment 1 and the  lowest of the Experiment 2 variances (for "Row, Row, Row Your Boat")  confirms that the variances are significantly different [F (1,7) =  24.83; p<.01], using Leven's test for homogenity of variances, and Satterthwaite's correction for unequal n, (Snedecor & Cochran, 1989).

Figure 3 shows the variability of tempos expressed as  per cent deviation from each song's mean tempo. When compared with  Figure 2, the greater variability in tempos is easy to see. This  replicates Halpern's (1988) finding that production variability on these  types of songs is large. Furthermore, the fast and slow performances  of each of the three songs showed that there is indeed a large range  over which people can produce familiar songs with lyrics. The song  "Happy Birthday" exhibited maxima and minima of 421 and 48 bpm  (with the mean across subjects being 284 and 76 bpm). "We Wish You  A Merry Christmas" exhibited maxima and minima of 129 and 22 (with  means of 102 and 36). "Row, Row, Row Your Boat" exhibited maxima  and minima of 280 and 41 (with means of 226 and 72). This large  deviation in "normal" speeds, coupled with the large range of possible  speeds, confirms that a given song can be sung across a very broad  range of tempos, and that lyric or articulatory constraints are  probably not playing a role in our Experiment 1 subjects' accurate memory for tempo.

DISCUSSION.  In Experiment 2, we replicated Halpern's (1988) finding that  the variability in tempos for popular songs that lack a tempo standard  is in the 10-20% range, well exceeding the variability of our  Experiment 1 subjects. The accurate performance in Experiment 1 does not seem to have been due to constraints imposed by lyrics. 

GENERAL DISCUSSION

Songs contain both pitch and tempo information during  their performance. What can we say about the mental representation  of songs in the brain? Drake and Botte (1993) argued for the  existence of a single brain mechanism that judges the tempo of  sequences (not merely the durations of intervallic events).  Judgment of tempos might be controlled by a central timing  mechanism located in the cerebellum (Helmuth & Ivry, 1995), the  operation of which is based on oscillatory processes (Ivry &  Hazeltine, 1995). Such an "internal clock" may not keep perfect time,  but be subject to 1/f noise (Gilden, Thornton, & Mallon, 1995). But  pitch perception seems to occur in brain systems separate from  time perception, beginning with frequency selective cells in the  cochlea, all the way through to the auditory cortex (Moore &  Glasberg, 1986). So it would seem that the perception of pitch and  tempo is handled by different systems. 

But memory for songs may somehow combine or link pitch  and tempo representations. Our intuition - despite the finding that there was not a statistically significant correlation between tempo  memory and pitch memory - is that the entire spectral-temporal  profile of a song is encoded in memory in some fashion and that  repeated listenings strengthen the trace. Pitt (1995) provided  evidence for a central representation of instrumental timbre, which  supports the notion that memory preserves a complex, spectral- temporal image.  


Figure 3. Distribution of tempos in Experiment 2.   (a) Happy Birthday, SD = 20%.  (b) We Wish You A Merry Christmas, SD = 17%. (c) Row,  Row, Row Your Boat, SD =11%.


In any event, it seems increasingly clear that human  memory encodes both the abstract and the relative information contained in musical pieces, and that people are able to access  whichever is required based on the given task. This supports  previous theoretical predictions that memory does encode absolute  features of the original stimulus, along with abstract relations  (Bower, 1967; Hintzman, 1986). This would also account for  peoples' ability to easily recognize songs in transposition, and for  our findings of being able to reproduce a particular absolute feature.  Premack (1978) offers an account of the relation between abstract  and absolute memory, suggesting that abstraction is only induced as  a response to an overburdened memory; that is, only absolute cues  are memorized until memory becomes taxed, and then the organism  forms an abstract rule. 

Some colleagues have wondered if our results are merely  the effect of "overlearning" of the stimuli, and suggest that our  findings are nothing more than a measure of how well the stimuli  were learned. But we agree with Palmer (S. E. Palmer, Personal Communication, October, 1994) who argues that increased learning  is just a matter of increasing the signal to noise ratio in memory  retrieval. To toss aside the present findings as "merely overlearning  " is to miss our point; to paraphrase Halpern (1992), we are  interested in the nature of what is encoded in memory when memory  is working well. Overlearned, or well-learned stimuli, provide the  cleanest measure of this. 

It is well established, at least anecdotally, that expert  musicians are capable of producing tempos from memory with great  accuracy. The present study found that even nonmusicians have an  accurate representation of tempo that they are able to reproduce. Considering this together with the results of  Levitin (1994) and other studies, we wonder (perhaps somewhat  facetiously) if what people encode in memory is the first bar of  what would be written music, including the key, time signature, and  metronome marking! This would be parsimonious, allowing the brain  to store only the temporal and melodic relations between tones, and  imposing temporal and pitch anchors only when needed. And it would  explain the ease with which people are able to identify changes in  pitch and timbre. But our subjective impression, based on  introspection, is that long term memory for music functions more  like what Attneave and Olson (1971) described as short term music  memory: 

    The circulating short-term memory of a tonal sequence  just heard, which is experienced as an auditory image extended in  real time (as a melody that 'runs through one's head'), typically  preserves the key or specific pitch values of the original. (p. 164). 

Attneave and Olson (1971) believed that in contrast, the long-term  memory trace is encoded based only on relations, or intervals. They  note that there is nothing about such a coding system that precludes  the additional storage of pitch or tempo information, "but it is  evident that normal individuals do not, in fact, preserve this kind of  information with any high degree of precision" (pp. 164-165). We  believe that the present study provides evidence against this point,  and that ordinary individuals do possess representations that are  more accurate than was previously believed. Furthermore, the  present study provides evidence that the two types of operations  present in musical listening - schematic reduction and unreducible  idiostructural information (Narmour, 1977) - appear to also be  present in long-term memory.    

REFERENCES 

Allen, G. D. (1975). Speech rhythm: Its relation to performance universals and articulatory timing. Journal of Phonetics, 3,  75-86. 

Attneave, F., & Olson, R. K. (1971). Pitch as a medium: A new  approach to psychophysical scaling. American Journal of  Psychology, 84, 147-166. 

Block, R. A. (Ed.) (1990). Cognitive models of psychological time.  Hillsdale, NJ: Erlbaum. 

Boltz, M. G. (1994). Changes in internal tempo and effects on the  learning and remmebering of event durations. Journal of  Experimental Psychology: Learning, Memory, and Cognition,  20, 1154-1171. 

Bower, G. H. (1967). A multicomponent theory of the memory trace.  In K. W. Spence & J. T. Spence (Eds.), The psychology of learning and motivation: Adances in research and theory (Vol.  1, pp. 229-325). New York: Academic Press.

 Braun, F. (1927). Untersuchungen öber das persðnliche tempo.  Archiv der gesamten Psychologie, 60, 317360. 

Brooks, L. R. (1978). Nonanalystic concept formation and memory for  instances. In E. Rosch & B. B. Lloyd (Eds.), Cognition and  categorization. Hillsdale, NJ: Erlbum. 

Collier, G. L. & Collier, J. L. (1994). An exploration of the use of  tempo in jazz. Music Perception, 11, 219-242. 

Collier, G. L. & Wright, C. E. (1995). Temporal rescaling of simple  and complex ratios in rhythmic tapping. Journal of  Experimental Psychology: Human Perception and  Performance,21, 602-627. 

Collyer, C. E., Broadbent, H. A., & Church, R. M. (1994). Preferred rates of repetitive tapping and catagorical time production.  Perception & Psychophysics, 55, 443-453.

Crowder, R. G., Serafine, M. L., & Repp, B. (1990). Physical interaction and association by contiguity in memory for the words and melodies of songs. Memory & Cognition, 18, 469-476.

 Dennett, D. C. (1991). Consciousness Explained. Boston: Little,  Brown. 

Desain, P. (1992). A (de)composable theory of rhythm perception.  Music perception, 9, 439-454 

Deutsch, D. (1991). The tritone paradox: An influence of language on  music perception. Music Perception, 8, 335-347. 

Dowling, W. J., & Bartlett, J. C. (1981). The importance of interval  information in long-term memory for melodies.  Psychomusicology, 1, 30-49. 

Drake, C., and Botte, M.-C. (1993). Tempo sensitivity in auditory  sequences: Evidence for a multiple-look  model. Perception & Psychophysics, 54, 277-286. 

Finke, R. A. (1985). Theories relating mental imagery to perception.  Psychological Bulletin, 98, 236-259. 

Fraisse, P. (1981). Rhythm and tempo. In D. Deutsch (Ed.), The  psychology of music (pp. 149-180). New York: Academic  Press. 

Friberg, A., & Sundberg, J. (1993). Perception of just noticeable  time displacement of a tone presented in a metrical  sequence at different tempos. In Kungl. tekniska  hoegskolan, Speech Transmission Laboratory [Eds.],  Quarterly progress and status report. Stockholm: Royal  Institute of Technology, Dept. of Speech Communication,  Speech Transmission Laboratory. 

Gilden, D. L., Thornton, T., Mallon, M. W. (1995). 1/f noise in human  cognition. Science, 272, March 24, 1995. 

Halpern, A. R. (1988). Perceived and imagined tempos of familiar songs. Music Perception, 6, 193-202.

Halpern, A. R. (1989). Memory for the absolute pitch of familiar  songs. Memory & Cognition, 17, 572-581. 

Halpern, A. R. (1992). Musical aspects of auditory imagery. In D.  Reisberg (Ed.), Auditory Imagery (pp. 1-28). Hillsdale, NJ:  Erlbaum. 

Hanson, H. M. (1959). Effects of discrimination training on stimulus  generalization. Journal of Experimental Psychology, 58,  331-334. 

Helmuth, L. L., & Ivry, R. B. (1995). When two hands are better than  one: Reduced timing variability during bimanual movements.  Journal of Experimental Psychology: Human Perception and  Performance, 22, 278-293. 

Hibi, S. (1983). Rhythm perception in repetitive sound sequence.  Journal of the Acoustical Society of Japan, 4, 83-95. 

Hintzman, D. L. (1986). "Schema abstraction" in a multipletrace  memory model. Psychological Review, 93, 411428. 

Hulse, S. H., Fowler, H., & Honig, W. K. (Eds.) (1978). Cognitive  processes in animal behavior. Hillsdale, N.J.: Erlbaum. 

Hulse, S. H., & Page, S. C. (1988). Toward a comparative psychology of  music perception. Music Perception, 5, 427-452. 

Hurly, T. A., Ratcliffe, L., & Weisman, R. (1990). Relative pitch recognition in white-throated sparrows, Zonotrichia  albicollis. Animal Behavior, 40, 176-181. 

Ivry, R. B., & Hazeltine, R. E. (1995). The perception and production of  temporal intervals across a range of durations: Evidence for  a common timing mechanism. Journal of Experimental  Psychology: Human Perception and Performance, 21, 3- 18. 

Jacoby, L. L. (1983). Remembering the data: Analyzing interaction  processes in reading. Journal of Verbal Learning and Verbal  Behavior, 22, 485-508. 

Jones, M. R., & Boltz, M. (1989). Dynamic attending and responses to  time. Psychological Review, 96, 459-491. 

Konishi, M., & Kenuk, A. S. (1975). Discrimination of noise spectra by  memory in the barn owl. Journal of Comparative  Physiology, 97, 55-58. 

Kosslyn, S. M. (1980). Image and mind. Cambridge, MA: Harvard.

Kuhn, T. L. (1974). Discrimination of modulated beat tempo by  professional musicians. Journal of Research in Music Education, 22, 270-277.

Kuhn, T. L. (1977). Effects of dynamics, halves of exercise, and trial  sequences on tempo accuracy. Journal of Research in Music Education, 25, 222-227. 

Levitin, D. J. (1994). Absolute memory for musical pitch: Evidence  from the production of learned melodies. Perception &  Psychophysics, 56, 414-423. 

Lockhead, G. R., & Byrd, R. (1981). Practically perfect pitch. Journal  of the Acoustical society of America, 70, 387-389. 

McClelland, J. L., & Rumelhart, D. E. (1986). A distributed model of  human learning and memory. In D. E. Rumelhart, J. L. McClelland, and the PDP Research Group (Eds.), Parallel  Distributed Processing, Volume 2: Psychological and  Biological Models (pp. 170-215). Cambridge, MA: MIT Press. 

Medin, D. L., & Shaffer, M. M. (1978). Context theory of classification  learning. Psychological Review, 85, 207-238. 

Michon, J. A., & Jackson, J. L. (Eds.) (1985). Time, mind, and behavior.  New York: Springer-Verlag. 

Monahan, C. B. (1993). Parallels between pitch and time and how  they go together. In T. J. Tighe & W. J. Dowling (Eds.),  Psychology and music: The understanding of melody and  rhythm. Hillsdale, NJ: Erlbaum. 

Moore, B. C. J., & Glasberg, B. R. (1986). The relationship between  frequency selectivity and frequency discrimination for subjects with unilateral and bilateral cochlear impairment.  In B. C. J. Moore & R. D. Patterson (Eds.), Auditory Frequency Selectivity (pp.  407-414). New York: Plenum Press. 

Narmour, E. (1977). Beyond Schenkerism: The need for alternatives  in music analysis. Chicago: University of Chicago Press. 

Perron, M. (1994). Checking tempo stability of MIDI sequencers.  Paper presented at the 97th Convention of the Audio  Engineering Society, November 10-13, 1994, San Francisco. 

Pitt, M. A. (1995). Evidence for a central representation of  instrumental timbre. Perception & Psychophysics, 57,  43-55. 

Posner, M. I., & Keele, S. W. (1970). Retention of abstract ideas.  Journal of Experimental Psychology, 83, 304-308. 

Povel, D. J. (1981). Interval representation of simple temporal  patterns. Journal of Experimental Psychology: Human  Perception and Performance, 7, 318. 

Povel, D. J., & Essens, P. (1985). Perception of temporal patterns.  Music Perception, 2, 411-440. 

Premack, D. (1978). On the abstractness of human concepts: Why it  would be difficult to talk to a pigeon. In S. H. Hulse, H. Fowler, & W. K. Honig (Eds.), Cognitive processes in animal  behavior (pp. 423-451). Hillsdale, NJ: Erlbaum. 

Reese, H. W. (1968). The perception of stimulus relations. New York:  Academic Press. 

Rosch, E. (1975). Cognitive representations of semantic categories.  Journal of Experimental Psychology: General, 104, 192-223. 

Serafine, M. L. (1979). A measure of meter conservation in music,  based on Piaget's theory. Genetic Psychology Monographs,  99, 185-229. 

Snedecor, G.W., & Cochran, W.G. (1989). Statistical methods (8th ed.). Ames, IA: Iowa State University Press.

Steedman, M. J. (1977). The perception of musical rhythm and metre.  Perception, 6, 555-570. 

Terhardt, E., & Seewan, M. (1983). Aural key identification and its  relationship to absolute pitch. Music Perception, 1, 63-83. 

Tooze, Z. J., Harrington, F. H., & Fentress, J. C. (1990). Individually  distinct vocalizations in timber wolves, canis lupus.  Animal Behavior, 40, 723-730. 

AUTHORS' NOTE 

This research was supported by a National Defense  Science and Engineering Graduate Fellowship to the first author, and  by NSF Research Grant BNS 85-11685 to R. N. Shepard. This report  was prepared in part while the first author was a Visiting Research  Fellow at the Center for Computer Research in Music and Acoustics,  Stanford University, Winter 1994-95. We are grateful to the  following for their generous contributions to this work: Jamshed  Bharucha, Chris Chafe, Jay Dowling, Andrea Halpern, Stephen Handel,  Douglas Hintzman, Jay Kadis, Carol Krumhansl, Max Mathews, Joanne  Miller, Caroline Palmer, John R. Pierce, Michael Posner, Bruno Repp,  Roger Shepard, Julius Smith, the researchers and staff of CCRMA,  and especially to Malcolm Slaney and the participants in the weekly  CCRMA Hearing Seminar. Any errors remaining in this work were  probably pointed out to us by these people, but we were too stubborn  to change them. Correspondence may be sent to D.J. Levitin, Behavioral Sciences Laboratory, Interval Research Corporation, 1801C Page mill Road, Palo Alto, CA 94304. (650) 842-6236 (email: levitin@interval.com)  P.R. Cook is now at the Department of Computer Science, Princeton  University. 

FOOTNOTES 

1. Only 88 data points are shown in Figures 1 and 2. We started with  46 subjects in Trial 1; three discontinued participation after Trial  1, and we eliminated the data of one Trial 2 subject. This Trial 2  subject asked to sing a Mozart Piano Sonata on Trial 2, complaining  that she didn't know any rock songs other than her Trial 1 choice  ("Hey Jude"). The experimenter let her sing the Piano Sonata. Later, during the analysis phase of the study, we realized that multiple recorded versions of such a piece exist, and virtually all are in the same key, so this subject's production was not excluded from the original pitch analysis reported in Levitin (1994). However, the  range of tempos over which such pieces are performed typically  varies, so this subject was excluded in the tempo analysis, on the  grounds that a single reference standard did not exist.

2. The low correlation between pitch memory and tempo memory could indicate that subjects have either independent storage or integrated storage of these two attributes, in the sense of the terms proposed by Crowder, Serafine, and Repp (1990). In independent storage, memory for one element is uninfluenced by the other; in integrated storage, the elements are related in memory in such a way that one component is better recognized in the presence of the other than in its absence. The present data do not provide direct evidence for choosing between these two possibilities, but it is our intuition that tempo and pitch are best characterized by an integrated storage account. Some subjects are able to recall both attributes with great accuracy (one attribute might indeed aid the accurate recall of the other) while other subjects showed no such benefit.