The Roles of Absolute Pitch and Timbre in Plink Perception

This research was approved by the St. Mary’s College of Maryland (SMCM) IRB before data collection began. We recruited 146 participants through the Psychology Department’s research participation pool but we lost one participant’s task data due to a computer crash. Due to a logistical error, we failed to collect questionnaire data from one other participant before their debriefing. Of the 145 participants for which we collected questionnaire data, they included 98 (67.59%) female participants, 46 (31.72%) male participants, and one participant (0.69%) who did not report their sex. Participant ages ranged from 17 to 24 (M = 18.81, SD = 1.16). The three participants who were under 18 provided parental consent prior to participation. Five participants reported a country of birth other than the United States (Australia, Cameroon, Germany, Greece, and Italy) and one participant did not report their country of birth. One hundred thirty-nine participants (95.86%) reported English as their native language, two participants (1.38%) reported Spanish, one participant (0.69%) reported Greek, one participant (0.69%) reported French, one participant (0.69%) reported Tagalog, and one participant (0.69%) did not identify their native language. All participants reported English to be their most comfortable spoken language at the point of participation, except for one participant who did not answer the question. Of the 144 participants who reported their confidence in having absolute pitch, eighty-four participants (58.33%) reported “definitely not,” thirty-one (21.53%) reported the value between “definitely not” and “perhaps a little,” twenty-four (16.67%) reported “perhaps a little,” four participants (2.78%) reported between “perhaps and little” and “yes, absolutely,” and one participant (0.69%) reported “yes, absolutely.” Participants reported listening to, on average, 32.72 (n = 134, SD = 29.92) hours of music per week (12 participants did not provide a response or provided impossible values). Participants were pseudorandomly assigned to one of four auditory conditions: original (n = 43), mistuned (n = 26), reversed (n = 30), or shuffled (n = 46).

It was important for us to choose recognizable music so that we could minimize the possibility of floor effects while comparing performance across acoustic conditions. Based on findings that the music individuals are exposed to during adolescence and early adulthood maintains a salient position in memory (e.g., Krumhansl & Zupnick, 2013), we selected highly recognizable Pop songs that charted during the time individuals within our population of interest (undergraduate students) attended middle school, high school, and college. We accordingly set our song release date timeframe from 2011 to 2018, targeting the Pop music we expected to have been playing across mainstream media during these years. First, we identified songs through an objective comparison of song popularity. We created an initial list of songs from multiple lists including “Billboard: Best Catchy Songs,” “Billboard: Top Ten Summer Songs of The Year,” and Grammy Awards Major Nominee of the Year. Next, directed research students, including author RNF-T, narrowed the initial song list to 50 hit songs (listed in Table 1) that were expected to be highly recognizable to students in the sampling population and we downloaded the songs from Amazon.com as .mp3 files. We converted each song to mono with Praat (Boersma & Weenink, 2018). Next, we extracted plinks from the original songs (example stimuli appear in Figure 1). Thiesen and colleagues (2019) pointed out that the position of the excerpt within the song can have dramatic effects on its ability to activate a listener’s memory. For example, Krumhansl (2010) created 300 and 400 ms plinks from chorus and non-chorus sections of her songs and found that participants recalled artist and song names more accurately in the 400 ms chorus condition (Thiesen et al., 2020, confirmed the chorus advantage). We imposed an additional rule to enhance overall plink recall: our plinks included spoken or sung vocal content. Krumhansl found that recall was better for plinks that contained word content from the song title than plinks that did not include such information (although Schellenberg et al., 1999, did not find a matching advantage for plinks with vocal content). Our plinks included vocal content but the vocal content never contained any part of the song title to reduce the possibility that semantic memory could drive memory effects across songs.

We created mistuned copies of each plink in Audacity (Audacity Team, 2018) by applying the “Change Pitch” function (semitones: -2, 2; with “high quality stretching” enabled, which uses a sub-band sinusoidal modeling synthesis algorithm to ensure that signal duration is preserved during pitch shifting). We originally planned to present sharp and flat versions that were mistuned by +/- 2 semitones (ST) based on Schellenberg and Trehub’s (2003) finding that individuals can accurately discriminate between original and 2 ST mistuned excerpts of familiar songs. However, to reduce the number of stimulus combinations in our study, we ultimately decided to only use the flat-tuned versions. Henceforth we refer to the -2 ST stimuli as mistuned.

Next, we created our timbre-modified plinks. We created reversed plinks by reversing the original plinks in Praat. The reversed and original plinks contained identical average spectral and pitch information but dissimilar fine-grained spectrotemporal information (Schellenberg et al., 1999). To create the shuffled plinks, we first considered procedures employed by other researchers. Levitin and Menon (2005) used scrambled classical music stimuli in an fMRI investigation of the brain bases of fine-grained, local feature processing in music perception. They scrambled their music stimuli by dividing long, 23 s excerpts into 250–350 ms segments and then randomly concatenated the segments with crossfade. Bigand and colleagues (2009) examined the role of musical feature stability for familiar and unfamiliar music by extracting and concatenating nonadjacent segments into a new signal. In their work, segments varied in duration between 250–850 ms (with fade) for a total signal duration of 11–26 s, depending on the piece. In contrast to these two studies and prior plink research, we presented listeners with shuffled 400 ms plinks. We wrote a Praat script to batch transform the original plinks into temporally shuffled versions according to the following algorithm (see Figure 2). For each original plink, the script divided the excerpt into 20 ms segments (there were 20 segments per 400 ms plink), applied 1 ms fade in/out to each segment (to avoid acoustic artifacts when the segments are combined), concatenated the segments in random order, and saved the shuffled plink. As can be seen in Figure 1, while the original and shuffled plink waveforms were dissimilar, their spectra were very similar (within each song). This indicates that our shuffle manipulation successfully disrupted local, dynamic spectral content while preserving global, average spectral content.

In total we created 50 plinks x 4 conditions (original, mistuned, reversed, shuffled) = 200 unique plinks. We created two additional stimuli. First, we created a plink from the song “Let It Go” by Idina Menzel to serve as a stimulus for practice trials only. The practice plink matched the characteristics of the plinks in the original condition except that it contained sound from two words of the title (“Let it”) to further enhance recognition. Second, we used Praat to create a 400 ms segment of random Gaussian noise to be used as a control stimulus. The control stimulus served to reduce the possibility of participants guessing the purpose of the study and to enable us to assess participant attention during the experiment (participants should not recognize or recall songs from the control stimulus). By including a stimulus that should not be recognized as a song, we were able to confirm the validity of our recognition measure by ruling out possible misuse of the “yes” response. Finally, we wrote a Praat script to normalize all stimuli (scale RMS intensities to 70 dB SPL; add 10 ms onset and offset ramps to eliminate sound presentation artifacts), and we listened to the stimuli to confirm the sound properties of each plink. Stimuli are available at https://osf.io/ma589/.

We measured plink memory in two ways: recognition and recall. According to the dual-process theory of memory, recognition (identifying familiarity with a stimulus) is a separate process from recall (recollecting the relevant semantic qualities of a stimulus) and they differ in time course, reliability, and neural activity (Hintzman et al., 1998; Mandler, 2008). First, we defined recognition as a “Y” keypress after excerpt exposure signaling familiarity to the stimulus. We chose to use a single keypress for recognition so that we could measure participants’ recognition RTs. When we designed the study, we reasoned that a latency measure might be sensitive to the perceptual processes responsible for plink memory across conditions (Hintzman et al., 1998). Thus, for example, we speculated that if one condition presented faster response times than another, it could indicate that the acoustic characteristics of that condition are better able to activate song memory. The disadvantage of the recognition measure is that there is no way to verify response accuracy. For example, a “Y” keypress could indicate legitimate plink recognition, a recognition error (identifying a different song or artist), or a response error (e.g., accidental keypress). We supplemented our recognition measure with another measure of memory. We defined recall as recollection of the correct song title and/or artist. The major advantage of this measure is that we can verify participants’ recall accuracy because we know the song and artist names for the plinks. The disadvantage of the recall measure is that, due to differences in participants’ typing speed, it cannot be used to infer anything about the time course of perceptual processing (thus we did not measure recall response time). We separately coded recall accuracy for song, artist, and both because song and artist recall accuracy can vary independently (e.g., if a listener knows the song name but not the artist). Taken together, our measures of recognition, RT, and recall provide a robust assessment of plink memory.

Participants completed three questionnaires. The general information questionnaire served to gather participants’ demographic information including age, gender, handedness, and hearing and spoken language pathology. We also included items to assess participants’ subjective comfort during the task and the amount of effort they exuded in completing it. The language background questionnaire assessed participants’ place of birth, native language, additional spoken languages, family language history, and, for nonnative talkers of English, age of English language acquisition and current comfort level in speaking their languages. The music background questionnaire assessed participants’ experiences with music including items related to their experiences with formal and informal music education. It asked participants to rate their own singing capabilities and their enjoyment and frequency of singing. The questionnaire also inquired about participants’ vocal and instrumental training and performance experiences. Finally, following a printed, traditional definition of absolute pitch, it asked participants to indicate if they have absolute pitch on a five-point scale (1 = definitely not; 5 = definitely yes).

Participants signed up for a 45-minute research session, during which up to five students could participate simultaneously, via an online participation platform. Participants received research participation credits to apply toward a relevant course grade. Data collection occurred in a room containing multiple computer stations that were divided with privacy partitions. Each station included a Windows PC with widescreen monitor, keyboard, mouse, and a pair of Audio-Technica ATH-M30 circumaural headphones. Following a review of the consent form, the researcher instructed participants to wear their headsets and begin the auditory portion of the experiment (described below). Participants completed the experiment at their own pace. They were asked to silently alert the researcher after they completed the auditory trials so that the researcher could provide a questionnaire packet for silent completion at the workstation. Upon completion, participants were debriefed outside of the experimental room to avoid disturbing other participants.

We programmed stimulus presentation and response collection with PsychoPy (version 1.90.2, Peirce et al., 2019). Throughout the auditory trials, the computer screen displayed a black background and light gray font. The PsychoPy script was loaded at each workstation prior to participants’ arrival. The initial screen instructed participants to wait for instructions from the experimenter, who directed participants to press “Enter” after completing informed consent. The next screens reminded participants to wear headphones and displayed instructions including a general description of the task and response options for recognition and recall of song and artist names. After reading the instructions, participants pressed “Enter” to continue to the practice trials (described below).

Each auditory trial utilized the following format. A blank screen displayed for 600 ms followed by on-screen text “Do you know this song?” At 1 s after the start of the trial, the 400 ms plink played three times with 1 s interstimulus intervals. We chose three exposures based on our concern that a single exposure was not sufficient if a participant did not attend to the beginning of the trial. Furthermore, because we were interested in their timed responses, we chose not to allow the participants to control stimulus repetition. At the onset of the third auditory presentation, on-screen text, “press Y (yes) or N (no),” appeared to remind participants about their song recognition response options. The computer was programmed not to accept keypresses until the beginning of the third presentation when the response instructions appeared. At that point, the response timer began so that we could measure recognition response time. If the participant typed “N,” the next trial began. If the participant typed “Y,” a pop-up window appeared for the participant to type the name of the song and artist and to rate their confidence in their song and/or artist guess by selecting a number between 0 (no confidence) and 5 (maximum confidence). Participants were not required to enter any responses at this window and they were informed during the instructions that spelling did not count. When either the “OK” or “Cancel” button was clicked, the popup window disappeared and the next trial began.¹

The practice trials served to acclimatize participants to the brief duration of the plinks and to enable them to practice responding with either a “Y” (yes) or “N” (no) keypress to indicate their recognition of the plink. The practice trials followed a similar format to the typical trials except that additional information appeared to the participant during practice. First, an instruction screen described the structure of the two practice trials. Next, the on-screen text for the first practice trial instructed the participant to respond “N” even if they recognized the song (the computer would not accept any other keypress). Following this “N” keypress, the second practice trial began. This practice trial explicitly displayed the name of the song and artist (Let It Go, Idina Menzel) and informed participants to press “Y” to continue (the computer did not accept any other keypress). Following this “Y” keypress, participants viewed the popup window and practiced entering the artist name, song name, and their confidence. Following the second practice trial, a final instruction screen stated, “The practice trials are over. Now you are ready for the real trials. Do your best to identify the songs. No matter what, please respond as accurately as you can.” Next, participants completed 75 randomly ordered trials including 50 plinks and 25 noise control trials. Because we utilized a between-subjects design, each participant heard plinks from just one of the four conditions (original, mistuned, reversed, or shuffled). Upon completion of the main phase of the experiment, a screen instructed participants to tell the experimenter they are finished so that they could begin the questionnaires.

We chose to utilize nonparametric inferential analyses due to the large, non-normal variation in data across participants and songs (Figure 3; recognition and recall performance is positively skewed across conditions). In other words, some participants recognized many plinks, and some participants did not recognize any. Likewise, some plinks were highly recognizable whereas others were never recognized. We report appropriate measures of central tendency (median) and variation (interquartile range; IQR) with each of the nonparametric inferential statistics and we report means and standard deviations in the tables. We conducted the inferential analyses in two ways. The listener analyses treated participant as a random factor to determine how condition affects plink memory across listeners. We conducted listener analyses with between-subjects Kruskal-Wallis tests and eta squared effect sizes (Cohen, 2013) followed by Dunn-Bonferroni post hoc pairwise comparisons (these include adjusted p values to reduce the possibility of false positives). The song analyses treated song as a random factor to determine how condition affects plink memory across songs. We conducted song analyses with repeated measures Friedman tests and Kendall’s coefficient of concordance (W; Cohen, 2013) as effect size, followed by Dunn-Bonferroni post hoc pairwise comparisons. The inferential results are occasionally presented with various sample size values due to missing data (e.g., some participants did not report recall confidence and those data are omitted from their respective analysis). For both the song and listener analyses of recall data, the input data comprises responses we coded as correct and exchanged. We omitted incorrect, song word error, and partially correct responses (see Table 2; for more detailed information on song response coding, see the Supplementary Materials accompanying the online version of this paper at mp.ucpress.edu). For this reason, our results are more conservative than they would have been had we included the partially correct responses. We reasoned that our predictions would be partially supported if they were upheld in either analysis and strongly supported if they were satisfied in both analyses. Due to the similarity between listener and song results, only listener results have been included in the main text. Please see the Supplementary Materials for song analysis results. Data files are available at https://osf.io/ma589/.

The Kruskal-Wallis test indicated a statistically significant difference in recognition between the four conditions, χ²(3, n = 145) = 67.75, p < .001, η² = .46. Dunn-Bonferroni post hoc tests revealed that the proportion of recognized plinks was higher in the original condition (Md = .32; IQR = .24) than the mistuned (Md = .14, IQR = .15; p = .001), reversed (Md = .14, IQR = .16; p < .001), and shuffled (Md = .10, IQR = .10; p < .001) conditions. Recognition was lower for shuffled plinks than mistuned (p = .006) and reversed (p = .010) plinks. There was no significant difference in recognition between the reversed and mistuned conditions (p = 1) (see Figure 3A).

The Kruskal-Wallis test revealed a statistically significant difference in recognition RT between the four conditions, χ²(3, n = 134) = 9.12, p = .028, η² = .05. The post hoc comparison indicated that response time was faster in the original condition (Md = 1.64 s, IQR = 1.11) than the shuffled condition (Md = 2.53, IQR = 2.10; p = .050). There were no other differences in RTs between the original, mistuned (Md = 2.61, IQR = 1.78), reversed (Md = 2.49, IQR = 3.16), and shuffled condition: original-mistuned (p = .357); original-reversed (p = .107); mistuned-reversed, mistuned-shuffled, and reversed-shuffled (p = 1) (see Figure 3B).

The Kruskal-Wallis test revealed a statistically significant difference in artist recall between the four conditions, χ²(3, n = 145) = 63.17, p < .001, η² = .43. The proportion of artists recalled was higher in the original condition (Md = .22, IQR = .20) than the mistuned (Md = .06, IQR = .07; p < .001), reversed (Md = .08, IQR = .09; p = .001), and shuffled (Md = .04, IQR = .05; p < .001) conditions. Artist recall was higher for reversed than shuffled plinks (p = .009), but the difference between mistuned and shuffled conditions was not statistically significant (p = .256). The mistuned and reversed conditions did not differ (p = 1) (see Figure 3A for song, artist, and both recall across the four conditions).

The Kruskal-Wallis test indicated a statistically significant difference in song recall between the four conditions, χ²(3, n = 145) = 75.25, p < .001, η² = .51. Post hoc comparisons revealed that the proportion of songs recalled was higher in the original (Md = .17, IQR = .16) condition than the mistuned (Md = .04, IQR = .06; p < .001), reversed (Md = .04, IQR = .06; p < .001), and shuffled (Md = .02, IQR = .04; p < .001) conditions. Mistuned song recall was higher than shuffled (p = .010) but not reversed (p = 1) conditions. Song recall between reversed and shuffled conditions did not statistically differ (p = .302).

The Kruskal-Wallis test indicated a statistically significant difference in recall of both song and artist between the four conditions, χ²(3, n = 145) = 59.03, p < .001, η² = .40. Recall of both song and artist was higher in the original condition (Md = .14, IQR = .16) than in the mistuned (Md = .04, IQR = .05; p = .001), reversed (Md = .02, IQR = .05; p < .001), and shuffled (Md = .02, IQR = .04; p < .001) conditions. Recall was higher for mistuned than shuffled plinks (p = .045) whereas the difference between reversed and shuffled was not statistically significant (p = .450). Recall did not differ between mistuned and reversed conditions (p = 1).

The Kruskal-Wallis test indicated a statistically significant difference in recall confidence between the four conditions, χ²(3, n = 125) = 37.31, p < .001, η² = .28. Confidence was statistically significantly higher in the original condition (Md = 3.29, as measured on a scale between 0 and 5, IQR = 1.04) than the mistuned (Md = 2.52, IQR = 1.79; p = .006), reversed (Md = 2.00, IQR = 1.55; p < .001), and shuffled (Md = 1.17, IQR = 1.46; p < .001) conditions. The difference between mistuned and reversed conditions did not differ (p = 1), nor did mistuned-shuffled (p = .171) or reversed-shuffled (p = .569) conditions (see Figure 3C).

Our sample was able to recognize and recall shuffled plinks at an above-zero level of performance, but performance in that condition was lowest overall. Based on an anonymous reviewer’s suggestion, we performed a series of one-sample Wilcoxon signed rank tests to test the null hypothesis that the population medians of recognition and recall are zero for shuffled plinks. The results revealed above-zero levels of recognition, z = 5.25, p < .001, r = .77, and recall of artist, z = 4.66, p < .001, r = .69, recall of song, z = 4.19, p < .001, r = .62, and recall of both artist and song, z = 4.02, p < .001, r = .59.

As we noted in the Method, one disadvantage for recognition measures of memory is that accuracy of responses cannot be determined because we do not know participants’ prior exposure to the songs in the stimulus set. But, if participants are truthfully reporting their recognition of each plink, then recognition should positively correlate with recall. Thus, we investigated the validity of our song recognition measure by calculating Spearman’s rank-order correlation between participants’ mean proportion of plinks recognized and mean proportion of plinks with both song and artist correctly recalled, r_s(143) = .79, p < .001. Though we report the overall correlation here, the correlations between recognition and recall were positive and statistically significant in all four conditions.

Given that we selected popular songs for our stimulus set, we wondered whether our participants’ listening habits were related to their performance. We correlated plink recall with participants’ self-reported weekly music listening (hours). There was a significant positive correlation between music listening and both artist recall, r_s(121) = .21, p = .021, and recognition r_s(119) = .23, p = .013. The correlations between weekly listening and song recall, r_s(119) = .15, p = .090, or both recall, r_s(119) = .16, p = .070, were not significant. We also correlated participants’ self-reported absolute pitch ratings and their recall performance, but the correlations were not statistically significant (all p values > .382).

Krumhansl (2010) asked her participants to rate their recognition of each plink on a Likert scale and the recognition ratings provided in her study serve a similar function as the confidence responses that we collected. If our confidence ratings correlate with recall ratings, then that would validate the use of recognition rating scales as a singular measure of plink memory. This could be a tempting strategy because it is much easier to collect plink recognition ratings than collecting and coding recall responses. The Spearman rank-order correlation between participants’ mean confidence ratings and mean proportion of plinks for which both song and artist were correctly recalled was positive and significant, r_s(123) = .50, p < .001. However, when we examined this correlation across all four conditions, it was only significant within the original condition, r_s(38) = .34, p = .030. This result suggests that recognition rating scales could serve as a valid predictor of recall, but only for unaltered plinks.

Does response time predict response accuracy? In other words, is there a relationship between response time and recognition or recall? We examined this question by first conducting a Spearman correlation between participants’ mean response time and mean proportion of plinks recognized. The result was negative and significant, r_s(132) = -.29, p = .001, indicating that lower overall response times were associated with higher recognition. However, when we examined this correlation within each condition, the only significant association we found was in the original condition, r_s(41) = -.38, p = .01 (all other p values > .10). Next, we examined the Spearman correlation between mean response time and mean proportion of plinks where both song and artist were correctly recalled. The result was r_s(132) = -.28, p = .001, indicating that the participants who entered their recognition responses more rapidly also recalled the plinks more accurately.

To test our prediction that artist recall would be better than song recall, we conducted a series of Wilcoxon signed rank tests on the listener data to compare artist and song recall within each condition. The results revealed that artist recall was statistically significantly better than song recall in the original (z = 3.27, n = 43, p = .001, r = .50), reversed (z = 4.10, n = 30, p < .001, r = .75), and shuffled (z = 3.57, n = 46, p < .001, r = .53) conditions, but not in the mistuned condition (z = .90, n = 26, p = .37, r = .18). We also examined the relationship between artist recall and song recall. We expected them to be positively correlated, and indeed, the Spearman correlation between participants’ mean proportions of artists and songs recalled was positive and statistically significant, r_s(143) = .87, p < .001 (this relationship held true within all four conditions; ps < .001). Based on condition means (see Table 3), we expected that artist recall would be higher than song recall overall, but we wondered whether the relative proportions of recalled artists and songs differed across conditions. To examine this possibility, we calculated the artist recall ratio for each participant as the ratio of recalled artists to recalled songs. If a participant recalled the same number of artists and songs, the artist recall ratio equals 1 (whereas larger values represent more artists than songs recalled and smaller values represent fewer artists than songs recalled). The artist recall ratio cannot be computed for participants who recalled zero artists and/or zero songs. Thus, 38/145 participants were removed from the overall analysis (original n = 0/43; mistuned n = 4/26; reversed n = 10/30; shuffled n = 24/46). Next, we submitted participants’ artist recall ratios to a between-subjects Kruskal-Wallis test that revealed a statistically significant difference in artist recall ratio between the four conditions, χ²(3, n = 107) = 12.72, p = .005, η² = .09 (original Md = 1.13, IQR = .73; mistuned Md = 1.00, IQR = .83; reversed Md = 2.00; IQR = 1.80; shuffled Md = 1.13, IQR = 1.00). Dunn-Bonferroni post hoc tests revealed that the artist recall ratio was significantly higher, thus indicating relatively more artists than songs recalled, in the reversed condition than in the original condition (p = .018), and in the reversed condition than the mistuned condition (p = .005), but there were no other significant differences: original-mistuned (p = 1); original-shuffled (p = 1); mistuned-shuffled (p = 1); reversed-shuffled (p = .242). In summary, participants recalled more artists than songs, but the artist recall ratio was highest in the reversed condition. We examine the implications of this and other results next.

We confirmed that listeners can identify very brief song excerpts presented with their original acoustic characteristics (Krumhansl, 2010) and in reverse (Schellenberg et al., 1999). We additionally discovered that listeners can identify mistuned plinks. However, the fact that listeners performed better in the original condition than in the mistuned condition suggests that listeners’ utilize their implicit absolute pitch memories (Levitin, 1994; Schellenberg & Trehub, 2003) to identify plinks. Our most surprising novel finding was that listeners were able to identify temporally shuffled plinks for which dynamic spectral content was substantially disrupted while average spectral content was spared. Taken together, our observations of plink identification for mistuned, reversed, and shuffled plinks strongly suggest that neither canonical tuning or temporal order of acoustic events is necessary for plink identification. However, the fact that listeners identified the original plinks more accurately than the other conditions, responded to original plinks most quickly, and reported the highest recall confidence ratings for original plinks strongly indicates that plink perception benefits when stimulus acoustic characteristics, including absolute pitch and timbre content, match long-term memory. But, it is important to note that listener performance varied dramatically. Our best performing participant, who was in the original condition, recognized 74% of the plinks and recalled the song and artist names for 64% of the plinks. In comparison, our second best performing participant, also in the original condition, recognized 52% of the plinks and recalled the song and artist name for 32% of the plinks. All but 10 (who were all in the shuffled condition) of our 145 participants recognized at least one plink while all but 15 recalled at least one song or artist. Notably, the best performing participant was the only person to self-report the highest score on our absolute pitch questionnaire item, and this observation would be expected based on the notion that absolute pitch information facilitates plink perception.

Perhaps the most fascinating question raised by our results is, how can listeners identify mistuned, reversed, and shuffled plinks given that their acoustic content has never been encoded in long-term memory? Of course, humans routinely perform such auditory feats. For example, listeners can perceive a sentence spoken by talkers with different vocal ranges, speaking rates, and within different acoustic environments. Likewise, listeners can identify a given melody performed by singers with different ranges, performance tempi, and so on. Given that most listeners develop their implicit relative pitch perception abilities in infancy and retain them through adulthood (Trainor, 2005), it is not surprising that melodies remain recognizable when presented in various keys. In general, shorter duration signals present less relative pitch information because melodic interval information builds over time. However, our 400 ms plinks are so brief that only a small number of rapidly produced events are present. If listeners do utilize relative pitch information to identify plinks, then relative pitch would contribute to plink recognition in both the original and mistuned conditions. Yet, relative pitch would probably not be as powerful a cue as timbre because timbre varies much more within and across plinks. Thus, relative pitch cues may contribute to plink identification, but they are not required for plink perception. What about absolute pitch cues? Schellenberg et al. (1999) found that listeners could match original, but not reversed (where absolute pitch cues are the same), 100 ms plinks to song names from a list. They interpreted this result to mean that “timbre is more important than absolute pitch for identifying popular recordings from very brief excerpts” (p. 645). However, this statement might be more appropriate for their stimuli, procedure, and results than ours because we did not observe any pairwise comparisons in which participants recognized or recalled reversed plinks any differently than mistuned plinks, although we did typically observe an advantage for mistuned and/or reversed plinks over shuffled plinks. The advantage for reversed plinks over shuffled plinks in our recognition and artist recall measures supports Schellenberg et al.’s assertion that listeners utilize local dynamic spectral cues. We extended their findings to show that listeners can use these dynamic cues even when they are presented in reverse (at least for our longer, 400 ms plinks). But, this finding does not imply that listeners do not use absolute pitch cues during plink perception. On the contrary, the fact that our participants identified original plinks more accurately than mistuned plinks supports the role of absolute pitch cues in plink perception. Just as pitch and timbre are entangled in the acoustic signal, it is likely true that pitch and timbre are entwined within listeners’ musical memories. For example, Siedenburg and McAdams (2018) found that the accuracy of timbre ordering judgments in novel sound sequences decreases for sequences that vary in pitch, suggesting that pitch variability influences timbre memory.

Is plink perception a bottom-up process that depends exclusively on the signal or is it a top-down process that integrates context from long-term memory? There may be more than one perceptual-memory process at work (e.g., Kahneman, 2011; Wixted, 2007) and these processes may leverage multiple sources of information such as acoustic content within the signal or stored in long-term memory. Humans can rapidly process acoustic information, even when that information is limited to brief durations, but one must not conflate plink duration with the time it takes to perceive a plink (Schellenberg et al., 1999). Identification might not occur until several seconds after plink presentation, and slower responses might be indicative of top-down processing. Recent research offers evidence for the bottom-up strategy. Jagiello et al. (2019) conducted EEG and pupillometry measurements during passive exposure to brief (750 ms) excerpts of familiar and unfamiliar music. They found that pupil dilation and EEG activity differed between familiar and unfamiliar excerpts within 400 ms of plink onset. We presented 400 ms plinks, and although recognition RT is not a direct measure of perceptual processes, our participants’ RTs varied widely.

Given the combination of our data and our anecdotal observations, we cannot claim that plink identification depends exclusively on the acoustic content of each plink. Instead, we suggest a broader interpretation in which each plink activates a vast array of encoded acoustic content in which a detailed memory trace associated with song identity might ultimately emerge. On the one hand, plink identification could be considered an example of perceptual categorization. By analogy, rapid visual presentation is sufficient to trigger object categorization (VanRullen & Thorpe, 2001). On the other hand, plink identification could be considered an example of musical prediction in which the plink triggers slower perceptual or cognitive processes that enable a listener to imagine unheard events (Keller, 2012; Pearce & Wiggins, 2012; Rohrmeier & Koelsch, 2012). A similar argument for slower processing of plink stimuli was proposed by Bigand et al. (2005). Based on their finding that listeners could distinguish 250 ms plinks extracted from high- and low-moving music, they argued that the induced emotions “are too refined to be simply derived from basic emotional properties of sound…these responses required a cognitive appraisal” (p. 435). However, the fact that plink identification depends on a vast store of music memories does not necessarily mean that identification is a top-down process.

The critical question of whether plink perception is influenced by music memory is a single example of a broader, ongoing debate regarding whether cognitive processes can influence perception (Firestone & Scholl, 2016). In the introduction, we suggested that it may be possible for a listener to match a plink’s transient acoustic events, whether in the original or reversed temporal order, to events stored in memory. Listeners could then utilize this low-level acoustic event information to build a more complex song representation (Layman & Dowling, 2018). Alternatively, listeners might match a plink’s average spectral content to music information in memory, but this latter process could proceed via a rapid, automatic bottom-up route or a slow, intentional, top-down route. Although we cannot decisively claim whether plink perception is bottom-up or top-down, we can comment on whether listeners utilize local and global information to identify plinks.

In their investigation of local and global feature processing in music perception, Bigand et al. (2009) demonstrated that listeners can distinguish between familiar and unfamiliar music excerpts (durations between 11–26 s) that have been acoustically scrambled (segment durations between 250–850 ms). In their discussion (p. 249), Bigand et al. cite another result from their lab in which listeners could distinguish familiar and unfamiliar music when the scrambled segment durations were as brief as 50 ms. Indeed, our results expand on Bigand et al.’s findings by showing that listeners can identify familiar plinks (400 ms duration) with scrambled segments (20 ms duration). We agree with Bigand et al. that the “color of sound” (p. 243), a timbral combination of voicing and instrumentation information, is perceptible in these signals. In the context of Bigand et al.’s scrambled excerpts that were approximately 50 times longer in duration than our plinks, it makes sense to consider each segment of the scrambled sequence as a local (e.g., 500 ms) representation of the global (e.g., 20 s) feature space. But in the context of our stimuli, each plink can be considered a local representation of the original, global (full-length) song. From that perspective, each 20 ms segment within a scrambled plink constitutes a hyperlocal representation of local feature space. The rapid presentation of randomly ordered segments in the shuffled plinks greatly diminishes a listener’s ability to identify specific musical events that occur in either local or hyperlocal feature space. Thus, we propose that listeners can identify shuffled plinks, albeit at low baseline levels of performance, by comparing average spectral content to long-term memory. After all, a shuffled plink contains the same average spectral content as the original plink (Figure 1), though their hyperlocal spectral features differ. Additional support for this average spectrum hypothesis comes from the fact that listeners generally recalled artists more often than songs and that the artist recall ratio was highest in the reversed condition where listeners recalled artists nearly twice as often as songs. We expected these results based on the fact that a given song sounds more similar to songs performed by the same artist than songs performed by another artist. Alternatively, the artist identification advantage might also occur because people can identify basic categories (i.e., artists) more readily than subordinate categories (i.e., songs; Rosch et al., 1976).

It is also possible that listeners can use both average and hyperlocal acoustic information to identify plinks, but additional research is needed to verify this. For example, if listeners can utilize hyperlocal features, then they probably will not need all 20 segments to identify a shuffled plink. Researchers could test this possibility by systematically reordering or removing segments (which will alter average spectral content) or by presenting listeners with a few hyperlocal acoustic events (event duration ≤ 20 ms). Here it is useful to note the variability in which listeners recognized and recalled plinks across our conditions (see Table 1). For example, 42% of listeners in the original condition and 22% of listeners in the shuffled condition recalled the song and artist names for Bodak Yellow (Cardi B). Yet, more listeners (72%) in the original condition and fewer (0%) in the shuffled condition recalled the song and artist names for Bad Blood (Taylor Swift). Patterns such as these may be due to random variability between listeners in each condition (e.g., listeners’ familiarity with each song) or the degree to which the dynamic spectral content in the shuffled plinks serendipitously matched the original plinks. Although we randomized segments in the shuffled plinks, it is possible that a string of temporally adjacent segments from the original plinks also appeared in proximity to each other within the shuffled version, and this might be enough to trigger recall. However, we examined this possibility and determined that the number of serendipitously adjacent segments did not correlate with shuffled plink recognition (nor did several other measures of temporal orderliness; see Supplementary Materials accompanying the online version of this paper at mp.ucpress.edu). This result supports our claim that listeners recognize shuffled plinks by extracting their average spectra.

Although we considered creating a unique, randomly shuffled plink for each trial, we ultimately decided to use a single shuffled version of each plink. Thus, it is possible that our shuffle manipulation algorithm did not completely eliminate dynamic spectral cues. Although dividing each 400 ms plink into 20 randomly ordered segments substantially distorted the dynamic spectral content, the spectral content within each hyperlocal segment was maintained. Future researchers may want to investigate the degree to which dynamic spectral content is crucial for plink perception by gradually and systematically disrupting the dynamic spectral content of each plink. Investigations of alternative timbre smearing manipulations could also be fruitful. Researchers could explore the influence of segment ordering, and potentially disconfirm the average spectrum hypothesis, by examining performance across several shuffled versions of any given original plink. Fine-grained acoustic comparisons of plinks across conditions could illuminate the importance of local and global spectral features in plink perception. And, although we believe that the non-zero recognition and recall proportions for the shuffled plinks provides strong evidence that listeners can use average spectral information to identify plinks, it is important to note that our results cannot provide proof because there are alternative explanations for listeners’ abilities to identify them. For example, it is possible that some participants may have heard previous participants discuss the stimuli. Or, given the task demands, subjects may have inferred that the shuffled plinks comprised highly modified Pop songs. This might have been sufficient to activate memory representations of popular (familiar and common) music artists and songs, which may explain occasional successful responses by chance alone.

As we noted earlier, our results bolster the line of work that has shown that listeners encode absolute acoustic properties in long-term memory. One question that emerges from this is, how many exposures are necessary for listeners to form these absolute memories? Agus et al. (2010) showed that listeners can implicitly learn and recognize random noise patterns of 4 s duration in as little as four presentations. Schellenberg et al. (2019) exposed participants to novel melodies in a lab setting. After a brief, ten-minute delay, they presented these melodies again, along with new melodies. Critically, they manipulated the pitch (Experiment 1) and tempo (Experiment 2) of the learned melodies to determine how these acoustic factors influence recall. Schellenberg et al. found that changes to key, regardless of magnitude, had a negative effect on melody recognition, as did tempo, though the effect was weak. Taken together, these results demonstrate that listeners encode pitch and tempo information within the first exposure to a piece of music, and that this information can be accessed for future tasks such as song recognition. By extension, we suspect that listeners could identify plinks from songs they have only heard and attended to once, but this claim remains to be tested in the laboratory.

An examination of the recognition and recall proportions across plinks (Table 1) and conditions (Table 3) might lead the reader to wonder how our observations compare to prior work. Participants in the original condition recalled the song and artist names for 15% of the plinks. In comparison, Krumhansl (2010, p. 343) and Thiesen et al. (2020, p. 341) reported identification rates of 26% and 7% for their 400 ms plinks. The variation in plink perception across studies suggests that stimulus and sample selection can have a substantial effect on the results. We chose songs that we expected to be maximally familiar to our sample whereas Thiesen et al. (2020) chose a mixture of recent and older songs based on stimuli from Krumhansl’s (2010) work. The distribution of recall rates across studies indicates that plink identification estimates are sensitive to participants’ familiarity with the songs used to create the plink stimuli. Relatedly, Spivack et al. (2019) presented listeners with song stimuli from top-charting popular music between 1940 to 2015. Their participants, who were similarly aged as ours, reported highest recognition for songs released between 2001 to 2015 and progressively lower recognition for songs released before 2000. Thus, considering that listeners are most familiar with the music they typically encounter, plink researchers should expect recall performance rates for any given stimulus set to decrease over time. This has consequences for future efforts to systematically replicate plink perception effects: the songs that are familiar to participants today are not the same songs that will be familiar to participants tomorrow (Greenfield, 2017). Instead of using the same stimuli as prior work, researchers may want to consider implementing a systematic stimulus selection rule such as the one employed by Spivack et al. In that case, some stimuli will be the same as prior work but recently released, chart-topping songs would be added for each new investigation.

There are few important limitations of our method and results. First, our findings should not be taken to indicate that some songs in our stimulus set are objectively more recognizable than others, although that may be true. As Krumhansl (2010) and Thiessen et al. (2019) pointed out, the place in the song from which a plink is extracted can have substantial effects on plink identification. We agree, and, although we do not know how plink perception would have been affected if we had extracted plinks from different locations within the song choruses, we believe that our general pattern of results would hold true. Another limitation of our work is that, unlike many RT experiments, we did not strongly emphasize speeded responses in our task instructions. The fact that we did not identify many RT differences, other than fastest responses for original plinks, could indicate that participants may have prioritized response accuracy over speed. It could also indicate that our RT measure was not as sensitive as it could have been (if, for example, we measured speeded response after a single plink presentation). Researchers who intend to investigate the time course by which listeners access implicit absolute pitch and timbre memories should be careful to design their studies to maximize RT validity and reliability. Furthermore, researchers should try to distinguish RT associated with recognition and recall. Additionally, we recommend that researchers who measure plink recall should carefully curate their stimulus list to eliminate songs with alternative titles, featured artists, and other characteristics that could complicate response coding. Participant factors might also have influenced performance. Although we selected our stimuli to be largely recognizable to the sample we recruited, it is likely that differences in participants’ musical tastes and listening habits caused variation in stimulus familiarity that our between-subjects design could not account for. Researchers could utilize within-subjects designs to investigate the role of participant familiarity on plink recognition and recall. Finally, we have no idea how participants would have performed if they were in a different experimental context that did not explicitly instruct them to recall song and artist names. It is reasonable to assume that some degree of recognition and recall accuracy is explained by spreading activation and priming (e.g., Bharucha, 1987). For example, it may be the case that participants who successfully recognize one plink will be more likely to recognize another due to remembered associations between songs and artists. Unfortunately, our stimulus set may have amplified this possibility because we included several songs performed by a single artist. However, it is also conceivable that such associations may have produced interference that ultimately reduced performance. To examine the possibility of plink priming effects, researchers could embed musical plinks with many other natural and artificial sounds or compare performance patterns between groups in which single or multiple songs from a given artist are included. Finally, though our results are compatible with the notion that relative pitch contributes to plink identification, our design was incapable of establishing such an effect. We recommend that researchers investigate the possible role of relative pitch cues by carefully manipulating pitch intervals within plinks. We suspect that relative pitch manipulations would reduce plink identification rates and thereby support the role of relative pitch in plink perception.

The plink paradigm emerged when researchers began to ask how quickly humans can identify music genres (Gjerdingen & Perrott, 2008) and which acoustic characteristics enable humans to recognize songs (Schellenberg et al., 1999). Despite the specificity of these original research questions, ongoing research has revealed that the plink paradigm is relevant to broad theoretical investigations about the rapidity of perceptual processes, the overlap between perception and cognition, the basic structure of the mind, and computational models of human performance. In a footnote, Krumhansl (2010) stated that Shazam (Apple, Inc., 2020), a popular music recognition and discovery application, could not identify her plinks. That is also true for our set of original, mistuned, reversed, and shuffled plinks. However, Shazam can reliably identify original music given a few seconds of input, and we also found that it could identify longer, mistuned excerpts. However, unlike human listeners, Shazam does not identify these longer excerpts when they are reversed or temporally shuffled (at least with our shuffling parameters). The differences in music identification performance between humans and computers underscore the different strategies that each employs. For example, Shazam appears to need more than one second of input and it relies on temporally anchored acoustic characteristics (Wang, 2003, 2006) such that major disruptions in the time domain decrease identification performance. Our results show that human listeners, on the other hand, are more flexible in their auditory search strategies. Of course, computer algorithms can be reprogrammed, and we hope that our results inspire the development of more robust auditory search engines. Humans have remarkable auditory perception abilities in part because they maintain vast stores of relative and absolute acoustic content in memory. In conclusion, we have shown that listeners utilize absolute acoustic cues to identify very brief musical sounds and that they utilize dynamic and average spectral content to perceive plinks. Given these observations, future researchers should be better positioned to investigate which sources of acoustic information are necessary or sufficient for plink perception.

Stimuli, data files, and other materials are available at https://osf.io/ma589/. We gratefully acknowledge Hannah Madden and Eduardo Elias for their assistance with data collection. We thank two anonymous reviewers for their feedback.