🧠 Auditory Working Memory Predicts Individual Differences in Absolute Pitch Learning

📋 Study Overview

🎯 Core Finding

📋 Study Design

Experiment 1 — Implicit Note Memory as the WM measure

Participants

• N=17 University of Chicago students (M=20.6 years, SD=2.6, range 18–26)
• No reported AP, variable musical experience (M=7.4 years, SD=4.8, range 0–14)
• Not specifically recruited for musical background

Working memory measure — Implicit Note Memory (INM) task

• Hear a 250-ms sine target tone → masked by 1000 ms of white noise → reproduce by adjusting a starting note (1–7 semitones above/below) using on-screen arrows — small arrows moved 33 cents, large arrows 66
• 64 trials total (4 target notes × 8 starting notes × 2 repetitions)
• Score = absolute deviation in 33-cent steps from target (lower = better WM precision)

Explicit pitch-labeling task

• Pretest: 60 trials — 12 piano notes (C4–B4) presented five times each, labelled by note name on a keyboard, no feedback. (The 180 belongs to the training phase: 3 blocks of 60.)
• Training: 180 piano notes (3 blocks × 60), same task with feedback after each trial
• Rote posttest: 60 notes (same pitches), no feedback
• Generalization posttest: 48 notes — 36 novel (untrained octave C5–B5 on piano; acoustic guitar in the trained octave and one octave below) plus 12 that repeated the trained stimuli
• Each trial separated by 1000 ms white noise + 2000 ms scrambled piano tones to prevent relative-pitch strategies

Retest (n=6)

• 6 of the 17 participants returned for delayed retest at M=184 days (~6 months) post-training
• Abridged rote posttest (48 trials) + full generalization posttest (48 trials), no feedback
• No reported rehearsal or retraining between sessions

Experiment 2 — Auditory n-back as the WM measure (non-musical)

Participants

• N=30, analyzed N=29. The excluded participant self-reported being "unsure" about having AP; the exclusion rested on scoring 97% after training while consistently misclassifying notes by exactly 2 semitones in the pretest
• UChicago students, staff, and community members (M=22.0 years, SD=4.2, range 18–32; 19 male)
• M=4.6 years music experience (SD=6.0, range 0–26). The Exp 2 sample started music significantly later than the Exp 1 sample (14.1 vs 9.6 years of age, p=0.02) and fewer of them had any formal instruction at all (18/29 vs 15/17, Barnard's exact test p=0.04)

Working memory measure — Auditory n-back (ANB)

• Spoken-letter stream, ISI 3000 ms; press "Target" if current letter matches the one n trials back
• Both 2-back and 3-back versions (in that order), 90 trials each (3 runs × 30 letters)
• 30-trial practice round before each test version
• Score = d-prime (signal detection theory) per task
• Critical: non-musical, non-pitch task — tests general auditory WM, not pitch-specific memory

Explicit pitch-labeling task

• Pretest: 48 trials (12 piano notes × 4 reps, randomized)
• Training: 120 trials (12 × 5 × 2 blocks) with audio + visual feedback
• Rote posttest: 60 trials (12 × 5), no feedback
• Generalization posttest: 48 notes beyond trained timbre + octave range (parallel to Exp 1)

📊 Key Results

Training Effects — Experiment 1 (N=17)

Both rote (t(16)=5.91, p<0.001) and generalization (t(16)=3.60, p=0.002) posttests were significantly above chance (8.33%, i.e. 1/12).

Training Effects — Experiment 2 (N=29)

Effects smaller than Exp 1. The paper's explanation is about spread, not level, and is hedged: "the compressed variability in musical instruction potentially explains why the overall absolute pitch learning in the current experiment was slightly lower." Both posttests were above chance (rote t(28)=4.21, p<0.001; generalization t(28)=2.98, p=0.006).

Read the generalization row carefully. Generalization performance was significantly worse than rote performance (t(28)=4.73, p<0.001) — that is the well-known 4.73. But the improvement of generalization over the pretest, which is what this column reports, was only marginal (t(28)=1.80, p=0.09). In Experiment 1 the same contrast was significant (t(16)=2.31, p<0.05). Since generalization performance is the dependent variable of every mediation model in the paper, this distinction matters.

In fairness to the authors: resting on the above-chance test rather than on this contrast, they summarise Experiment 2 as showing that "participants showed significant improvements in both rote and generalized learning as a function of training." We report the weaker contrast because it is the one the mediation models depend on — but that is us being stricter than the paper, not the paper hedging.

Working Memory Predicts AP Learning

Six-month Retention (Exp 1, n=6)

• Average delay: M=184 days (SD=22)
• Rote posttest dropped from ~50% (immediate) to ~38% (delayed); generalization from ~29% to ~24%
• Loss of ~11 percentage points from immediate to delayed. On generalization the drop was ~5 points (29.2%→24.0%) and the paper reports it as a null: "this loss, however, was not statistically significant [t(5) = 1.86, p = 0.12]" — a null the authors use in favour of retention
• Rote still significantly above chance (t(5)=5.80, p<0.01); generalization only marginally so (t(5)=2.13, p=0.08). The authors do not claim retention is established — they claim the opposite view, that all gains vanish without rehearsal, "may not be completely accurate"
• No participants reported actively rehearsing notes between sessions
• Caveat: small retest sample (n=6) limits strong conclusions about retention. The 6 who returned also had marginally more musical experience than those who did not (M=9.50 vs 5.32 years, p=0.10), which is part of why their immediate rote score (~51%) sits above the full sample's 36.2%

🧠 Theoretical Implications

Reframing the "Critical Period"

• Traditional view: AP requires exposure during a critical period (typically before age ~6); adults cannot acquire it
• This study: The age-of-onset effect is statistically mediated by auditory WM. Early training may help mainly because it strengthens domain-general auditory WM, not because of a hard temporal window for AP itself
• Quote (authors, pp. 106–107): "The current set of studies cannot directly address whether post-critical period adults can gain absolute pitch ability that is comparable to 'true' AP ability (as our training paradigm was only a single laboratory session, and no participant reached a level of performance that is typically seen in a 'true' AP population). Consequently, the present results cannot directly comment on the underlying mechanisms of the phenomenon of 'true' AP. However, our finding across two studies that auditory working memory can explain the success of non-AP possessors learning absolute pitch categories supports the notion that intermediate levels of absolute pitch ability (operationalized as significantly above chance, but significantly below the level of performance typically observed among 'true' AP possessors) might be best conceptualized as a domain-general perceptual learning task, rather than a specifically musical ability"

Two-Step Model of AP

Domain-General Perceptual Learning

• AP category learning is best framed as perceptual category learning (Goldstone, 1998), not a uniquely musical skill
• Working memory has been shown to predict success in many other category-learning tasks (DeCaro, Thomas, & Beilock, 2008; Lewandowsky et al., 2012)
• But AP does not fit either COVIS mould. The authors are explicit: absolute pitch category learning "does not fall within the realm of information-integration, as there is one salient perceptual dimension (pitch)"; nor is it a verbalizable rule-based structure. They say this is what made the result uncertain in advance — a null or even negative relationship with working memory would have been defensible, which is why the positive finding is informative

Comparison to Gervain et al. (2013)

• The valproate group in Gervain (2013) reached ~28.3% (5.09/18) accuracy after a full week of pharmacologically-assisted training
• In the current study, high-WM individuals (>1 SD above the mean) reached 43.8% rote / 30.5% generalization after a single laboratory session, no drug
• Authors' interpretation: critical-period framing for AP "perhaps... need not be applied" — what looks like a closed window may instead be the consequence of insufficient general auditory WM in average adults

🔗 Connection to Other Research

Theoretical Foundations

• Levitin (1994): the two-step characterisation of AP (pitch memory + labelling) is prior to this paper, which cites it as such alongside Ross, Gore & Marks (2005) and Zatorre (2003). What these experiments add is mapping auditory working memory onto the first step
• Ross, Olson & Gore (2003): the implicit note memory paradigm used here as the auditory WM probe. (The same first author later published the review Ross, Gore & Marks, 2005 — two different papers, often conflated)
• Deutsch & Dooley (2013): AP possessors have larger auditory digit spans than matched non-AP musicians. This paper explicitly refuses the direction: the finding "while correlational, suggests that individuals may develop what is conventionally known as AP because they have a high auditory WM capacity, though the reverse is also possible (i.e. individuals first gain AP and then improve their auditory WM)"

Direct Successor — Van Hedger 2019 (PLoS ONE)

• The 2019 follow-up uses high auditory WM as a participant-selection criterion — an explicit operational consequence of the 2015 finding
• 2 of 6 high-WM adults achieved genuine AP after 8 weeks of training, validating the predictive role identified here
• Without the 2015 mediation finding, the 2019 selection rule would be unmotivated; the two papers together form a single argumentative chain

Pharmacological Counterpoint — Gervain et al. (2013)

• Argues critical period can be reopened pharmacologically (valproate / HDAC inhibitor)
• The current paper offers an alternative explanation for Gervain's success: high-WM participants in a behavioral protocol can reach comparable performance without drugs
• Suggests valproate's apparent benefit may operate via general WM/attention rather than by strictly "reopening" a critical period. The chain the authors propose runs valproate → dopamine → working memory: valproate has been suggested to indirectly affect dopamine release in prefrontal cortex in rat brains (Ichikawa & Meltzer, 1999), and phasic dopamine is implicated in updating context information in dorsolateral prefrontal cortex. Their own caveat: these findings "do not conclusively show that auditory working memory changes as a function of valproate"

⚠️ Limitations & Caveats

Performance below "true" AP

• High-WM participants (>1 SD) reached 43.8% rote / 30.5% generalization; the highest mean seen in any subgroup was ~51% (immediate rote among the 6 retested). That is well below the "near-perfect accuracy" the paper attributes to AP possessors — a level the paper describes qualitatively and never puts a number on
• Authors are careful: this is "intermediate AP" or AP category learning, not full "true" AP
• Cannot directly determine whether longer training would close the gap

The prevalence problem — raised by the authors themselves

• "If the phenomenon of AP was largely accounted for by individual differences in auditory working memory capacity, then one might expect AP to be much more common than the often cited statistic of one in every 10,000 individuals in Western cultures." High auditory WM is not one in ten thousand — so WM cannot be the whole story
• Their suggested additional factor: early musical training prioritises relative pitch, and relative pitch category learning develops "at a direct cost to absolute pitch sensitivity" (Dye, Ramscar & Suh, 2011)
• They also grant that "the mechanisms of acquiring absolute pitch categories among a 'true' AP population may be entirely different" from what these experiments measured

Sample size and generalization

• Exp 1 N=17, Exp 2 N=29 — small to moderate by perceptual-learning standards
• Both samples drawn from UChicago community (volunteer/convenience)
• Retest sample (n=6) particularly small — retention claims are tentative

Single-session limit

• One laboratory session (~1 hour) cannot speak to the upper bound of trainable adult AP
• Different from Sakakibara (2014) or Wong (2025) protocols that span weeks/years

WM construct concerns

• Auditory n-back's construct validity as a WM measure has been questioned (Kane et al., 2007); recent work links it more to fluid intelligence than to WM per se
• Authors note INM and ANB share only ~25% variance, suggesting they tap different aspects of WM (quality vs. quantity)
• Could alternative WM measures (RSPAN, OSPAN, reverse digit span) replicate the mediation? Open question

Causal interpretation

• The mediation finding is correlational — cannot prove that improving WM would improve AP learning
• Possible reverse causality: people born with high auditory ability may have started music earlier because of that ability

🎯 Practical Implications

For Adult Learners

• Single-session gains are real: non-AP adults more than doubled their pretest accuracy in one session of feedback training — 13.7%→36.2% in Exp 1 (2.6×) and 10.9%→25.8% in Exp 2 (2.4×), against a chance level of 8.3%
• Individual differences matter: auditory WM is a meaningful predictor — high-WM individuals see substantially larger gains
• Musical background works through an indirect route: the effect of age of onset runs through auditory working memory, not alongside it. Mediation does not mean the background is irrelevant — age of onset correlates strongly with the WM measure itself (r=0.64 in Exp 1). The authors' formulation: early exposure "would not be necessary… but it might facilitate" AP learning by strengthening general auditory processes
• Some retention, on thin evidence: after ~6 months without rehearsal, rote performance was still above chance (n=6); generalization only marginally so (p=0.08). The authors' claim is modest — that the view "all gains are lost without rehearsal" may not be entirely right

For Researchers

• Selection design: screen participants by auditory WM before training — reduces variance and increases statistical power (this is what Van Hedger 2019 did)
• Mediation testing: always include WM as a covariate when reporting age-of-onset effects on AP learning
• Combined interventions: WM training (e.g., n-back training) plus AP training may compound — untested but theoretically motivated

For Music Educators

• Pre-screening for auditory WM may help identify students with the highest chance of acquiring AP through training
• The framing "critical period closed" may be misleading for adult students — recast as "general perceptual capacity" instead

📖 Methodology Details

Stimuli

• INM task (Exp 1): 250-ms sine waves, 4 target notes (F#4, G4, G#4, A4), 8 starting tones (D4, D#4, E4, F4 below; A#4, B4, C5, C#5 above), 33-cent step resolution
• Pitch-labeling task: Real instrumental notes sampled from Reason 4.0 software + Adobe Audition recordings, normalized to 75 dB SPL, 44.1 kHz
• Generalization stimuli (both Exp): 48 notes = 12 piano in the trained octave (C4–B4) + 12 piano one octave higher (C5–B5) + 12 acoustic guitar in the trained octave (C4–B4) + 12 acoustic guitar one octave lower (C3–B3). So the low octave was guitar, not piano — and 36 of the 48 were novel (new timbre, new octave, or both); the remaining 12 repeated the trained stimuli exactly
• Inter-trial masking: 1000 ms white noise + 2000 ms of 16 randomized scrambled piano notes (suppresses relative pitch use)

Apparatus

• Sennheiser HD280 studio monitor headphones
• 1280×1024 monitor, 75 Hz refresh rate
• INM run in Psychophysics Toolbox (MATLAB); explicit pitch-labeling run in E-Prime

Statistical Analysis

• Repeated-measures ANOVA for pretest vs rote vs generalization comparisons; Fisher's LSD post-hoc
• Generalized mixed-effects models (binomial link) with WM score and age of music onset as fixed effects, participant and stimulus note as random effects
• Mediation: Sobel test + bootstrapped indirect effects (10,000 samples; Preacher & Hayes 2008)
• One-sample t-tests for above-chance comparisons (chance = 1/12 = 8.33%)

🔬 Future Directions Identified by the Authors

• Extended training: Can high-WM adults reach "true" AP levels with longer protocols? (Answered partially by Van Hedger 2019 and Wong 2020/2025)
• WM training transfer: Would n-back or similar WM training programs improve AP trainability?
• Genuine AP populations: Test whether auditory WM also explains within-AP variability among "true" AP possessors
• Mechanism studies: Investigate whether valproate's apparent boost in Gervain (2013) operates via WM/attention enhancement rather than through critical-period reopening per se
• Construct refinement: Compare INM, n-back, RSPAN, OSPAN, reverse digit span as WM proxies for AP learning

💡 Key Takeaways

📚 Citation