AUDITORY LEARNING · MUSIC · PROSPECTIVE RESEARCH PROTOCOL
How does learning change the pleasure
of the same musical event?
Amaç Erdem | amace@bu.edu | ORCID 0009-0000-1711-0036
Abstract. The same musical event may feel different after different listening histories. We will combine participant-provided listening records
with Musical Intelligence (MI) features from verified recordings to estimate individual exposure profiles. Recognition and continuation tasks
will test these estimates. A controlled experiment will then manipulate learned expectations while keeping test audio unchanged and measure
pleasure. Separate comparisons will test whether history-informed MI adds predictive value beyond self-report, listening frequency and
acoustic baselines. Exposure is treated as evidence about listening experience, not a direct measurement of memory.
Academic edition r3 · 12 October 2026 · Source: v3.0, 7 October 2026. Proposed design; not preregistered; no new participant results.
Different learned histories
Identical test audio
Measure pleasure
P(A | cue X) = 1/2
P(A | cue X) = 1/8
Cue X
Target A
Does experience change
the response?
Assigned mappings · schematic illustration · no results shown
Fig. 1. The critical comparison holds the cue–target waveform constant across assigned histories.
1. The unresolved question
Research connects musical pleasure to surprise and uncertainty,
but reports different directions of their interaction [1, 2]. Learn-
ing can also create preferences for unfamiliar musical struc-
tures [3]. We ask how measured listening experience and exper-
imentally learned context relate to the response to an unchanged
musical event.
We will manipulate the probability of two target continua-
tions and the uncertainty of the cue that precedes them. Cue
meanings will be counterbalanced across listeners. Thus, an
identical cue–target recording can have different learned mean-
ings without changing its waveform.
2. Two distinct tests
H1. Contextual uncertainty changes the difference in pleasure
between less- and more-expected targets. The primary contrast
is a difference of differences, tested in both directions.
H2 · Secondary.
MI-derived exposure profiles improve
held-out prediction of familiarity, expectation and pleasure be-
yond self-report and listening-history baselines. H1 tests exper-
imentally assigned learning; H2 tests prediction from observed
histories. Neither establishes a unique neural mechanism.
3. What is held constant?
Within each musical family, target A is the same audio file wher-
ever it occurs; the same applies to B. Each target is heard equally
often overall. Four cues assign A and B probabilities of either
1/2 or 1/8, under low or high uncertainty. Three other continua-
tions preserve these distributions.
This separates conditional predictability from simple tar-
get familiarity. Cue acoustics and target identities are counter-
balanced, with sequence and repetition effects included in anal-
ysis. Specific cue–target pair familiarity remains linked to prob-
ability and limits interpretation.
4. Scope and participants
The planning sample is 192 adult non-musicians, plus a sepa-
rate pilot of 24. Twelve new musical families will be tested; each
listener learns one. English comprehension, headphone screen-
ing [6] and task practice are required.
Musical experience and reward sensitivity will be mea-
sured [4, 5]. History donation is optional; its absence does not
exclude a listener from H1. H2 uses the consenting subset with
usable records, with coverage and selection reported. Neither
dopamine nor brain activity is measured.
Listening history and the exposure profile
Identify what was heard
Participants may provide track records from Spotify, Apple Mu-
sic or another source. Keep title, artist and available recording
identifiers. Counts, dates and listening durations are used only
when supplied; a list of names cannot establish frequency or re-
cency.
Locate candidate recordings through catalogues or pub-
lic pages, including YouTube.
Confirm artist, version and
duration rather than accepting the first title match.
Record
live/cover/remaster differences and unresolved matches. Track-
level listening duration does not identify which moments were
heard.
Build a musical exposure profile
Analyse the verified recording using a frozen MI configuration.
R3 describes acoustic structure; T3 describes its temporal or-
ganisation. A separate comparison adds the six prespecified C3
outputs already shown in the architecture. Their names remain
computational labels.
Combine those features with the available history to estimate
exposure to musical structures. Prespecify aggregation, recency
weighting and missing-data rules in development data. Report
matched-track and listening-time coverage where possible. A
profile is an estimate to test, not a diagnosis or a reconstruction
of autobiographical memory.
5. Learning, listening and reporting
Before testing, consenting listeners provide Spotify, Apple Mu-
sic or equivalent track records. Names and artists locate candi-
date recordings; versions are verified before MI analysis. Fre-
quency and recency are used only when present. A title list sup-
ports a repertoire profile, not a timed exposure history.
An added recognition and continuation session validates the
frozen profile, using separate calibration and held-out items.
The controlled experiment retains its planned 50–60 minutes;
the pilot will determine the added history-task burden. Pleasure
and later secondary responses remain in separate blocks.
After consent, sound checks and practice, the session con-
tains 128 listening trials followed by 128 pleasure trials. Reso-
lution and replay follow in separate blocks.
Continuous pleasure · primary
“How much pleasure do you feel right now?”
Only this response is requested during the primary test. A
stationary, initialised slider is a valid response.
The response scale runs from 0 (none) to 100 (very strong).
Learning and the pleasure of identical music | Research proposal · 12 October 2026
1


a
What each cue predicts
b
The four target conditions
ASSIGNED DISTRIBUTIONS
CROSSED WITHIN EACH LISTENER
Cue
A
B
C
D
E
L1
low H
50%
12.5%
37.5%
0%
0%
L2
low H
12.5%
50%
0%
37.5%
0%
H1 high H
50%
12.5%
12.5%
12.5%
12.5%
H2 high H
12.5%
50%
12.5%
12.5%
12.5%
More expected
p = 1/2
Less expected
p = 1/8
Low uncertainty
1.41 bits
1 bit
32 target trials
3 bits
8 target trials
High uncertainty
2.00 bits
1 bit
32 target trials
3 bits
8 target trials
Cell values: target surprisal, −log₂ p
Matched marginal exposure
A = 40
B = 40
Rare targets remain rare in the test.
Fig. 2. Probability and uncertainty are separated by design. A and B are the focal targets; C–E are fillers. Across four cues, A and B each occur 40 times in
both exposure and test. Each test yields 80 target trials: 32 more-expected and 8 less-expected trials at each uncertainty level. Test frequencies preserve training
probabilities; cells are not artificially balanced by overplaying rare targets. Cue-to-distribution assignments vary across listeners.
a
One trial · one continuous report
PLEASURE ONLY
Musical cue
Target A or B
Silent response interval
0 s
2 s
3 s
4.5 s
7.5 s
9 s
Baseline
Delayed response window
−1 to 0 s from onset
1.5 to 4.5 s from onset
ΔPleasure = mean response −mean baseline
Audio-clock alignment · 20 Hz sampling
Fig. 3. An event-aligned outcome. Pleasure is sampled against the audio clock at 20 Hz. The fixed response window allows for reporting delay; it does not identify
the exact onset of an internal feeling. Interrupted trials and uninitialised responses are flagged.
Predicted continuation · learning check
“Out of 100 occasions, how often would each ending follow this
cue?”
Four cue-specific allocations before exposure, after exposure
and after the primary test. No outcome feedback.
Allocations across A–E sum to 100.
Perceived resolution · secondary
“Press when the musical idea feels settled.”
Sixteen later trials. Closure is kept distinct from pleasure;
non-response is not automatically failed attention.
The R key marks a moment.
Replay preference · secondary
“Which sequence would you like to hear again?”
Eight pairs; one choice is implemented at the end. Curiosity
and pleasure may make different predictions.
6. What Musical Intelligence can explain
The original audio architecture is retained. R3/T3 descriptors
and six existing C3 outputs form candidate representations. A
separate, proposed listener-profile layer combines these mea-
surements with observed listening history; it is not an imple-
mented memory mechanism.
Audio representation + listener history. Frozen MI re-
turns the same scores for the same recording. The proposed
profile layer adds track identity, repeat exposure, recency and
feature similarity where available. Its aggregation rules are fit-
ted in development data and frozen before held-out testing. Pro-
file scores are hypotheses about familiarity, not memories or au-
tobiographical diagnoses. The current MI comparison remains
offline.
7. Analysis, decisions and next steps
7.1 One primary behavioural test
For each target trial, we calculate the change in pleasure spec-
ified in Figure 3. A mixed-effects model estimates probability,
uncertainty and their interaction, accounting for listeners, musi-
cal families, recordings and trial order. H1 tests the interaction
at two-sided 𝛼= .05; effect sizes and 95% intervals are primary
reporting units.
Probability allocations test whether learning occurred.
Main-study listeners are not excluded for weak learning or low
pleasure. Failed learning limits interpretation of a null H1. Res-
olution and replay are secondary; neither establishes causal me-
diation.
7.2 Controlled-learning comparison
Four models predict the same pleasure outcome: A, R3/T3
acoustics and timing; B, A plus six MI scores; C, A plus a se-
quential cue–continuation learner; D, C plus the MI scores. B–A
tests MI’s audio representation; D–C tests its contribution be-
yond an explicit learning account (Fig. 5 in the original external
figure atlas).
The learning model receives the actual ordered exposure his-
tory. All models share matched acoustic access and fitting bud-
gets. Entire musical families and their listeners are withheld to-
gether; tuning uses development folds only. The primary pre-
dictive metric is held-out mean absolute error. Predictive gains
do not validate a named brain mechanism.
7.3 Pilot and registration gate
The pilot uses separate musical families to check learning, dis-
crimination, timing and burden. N = 192 is a planning figure, not
a power result. Simulation must support detection of a four-point
interaction on the 0–100 pleasure scale before the final sample is
registered. Stimuli, randomisation, exclusions, windows, code
and model versions are frozen before recruitment.
Learning and the pleasure of identical music | Research proposal · 12 October 2026
2


Full manuscript map
Shared selected inputs
F2 Prediction
F5 Emotion
F6 Reward
SPH  Spatiotemporal Prediction Hierarchy
PWUP  Precision-Weighted Uncertainty Processing
WMED  Working Memory-Entrainment Dissociation
UDP  Uncertainty-Driven Pleasure
SRP  Striatal Reward Prediction
RPEM  Reward Prediction Error in Music
G
J
K
L0
L0
L2
L2
L1
L1
SPH
SPH
PWUP
PWUP WMED
WMED
UDP
UDP
SRP
SRP
RPEM
RPEM
S1
S1
S2
S2
S3
S3
S4
S4
S5
S5
S6
S6
A
B
C
D
F
H
Audio
R³
T³
C³
9 groups · 97 features
Model readouts
Model input
44.1 kHz
mono
Consonance
7
Energy
5
Timbre
9
Change
4
Pitch/chroma
16
Rhythm/groove
10
Harmony
12
Timbre extended
20
Modulation
14
0
0.5
1
Excerpt mean
32 horizons
5.8 ms
Micro
Meso
Macro
Ultra
981 s
L0
L2
L1
24 operators
3 temporal rules
Features / cell
1
30
Past
Centred
Forward
89 mechanisms
941 outputs
131 beliefs
R
E
A
I
H
Core
App
Ant
F1 Sensory
F2 Prediction
F3 Attention
F4 Memory
F5 Emotion
F6 Reward
F7 Motor
F8 Learning
26 regions
4 proxies
A1/HG
STG
STS
IFG
dlPFC
vmPFC
OFC
ACC
SMA
PMC
AG
TP
VTA
NAcc
Caud
Amyg
Hipp
Put
MGB
Hypo
Ins
IC
AN
CN
SOC
PAG
DA
NE
OPI
5HT
Fig. 4. The implemented routes used in the model comparison. Coloured routes use the manuscript’s F2/F5/F6 palette; unused routes remain gray. Dashed
lines read R3 directly; solid lines pass through T3 or upstream mechanisms. The selected paths use 17 R3 features and 49 temporal keys. PWUP and WMED supply
internal values. The 28 September source audit is historical, static evidence. Counts are declarations; archived artwork values are not study results. Regional and
chemical outputs are outside the experiment.
Table 1. Prespecified MI candidates. Incremental prediction is tested jointly; named-mechanism interpretations remain exploratory.
SPH.P1.prediction_error
Audio-derived prediction-error candidate
UDP.P0.context_assessment
Audio-derived context candidate
UDP.P2.reward_computation
Pleasure-related candidate
SRP.P0.wanting
Approach-related candidate
SRP.P1.liking
Pleasure-related candidate
RPEM.current_rpe
Reward-error candidate, not measured RPE
7.4 Listening-history validation
Compare self-report alone, self-report plus available listening
counts/recency and artist/genre baselines, then the same model
plus an MI-derived exposure profile. Profile ablations use R3,
R3 + T3, and R3 + T3 + the six C3 candidates, with matched
fitting budgets. Freeze histories before testing; evaluation items
never tune the profile.
Primary history validation is recognition discrimination on
held-out familiar candidates and newly created, matched foils;
continuation prediction and pleasure are secondary outcomes.
”Absent from the export” does not mean unheard. Hold out lis-
teners and recording families; report matching coverage and sen-
sitivity to omitted tracks. Predictive improvement is not a causal
effect of past listening.
Learning and the pleasure of identical music | Research proposal · 12 October 2026
3


7.5 Recording identity and retention
Match title, artist, recording version, duration and identi-
fiers where available. Catalogues and public pages, including
YouTube, may locate candidates; uncertain, live, cover or remas-
tered substitutes are flagged. Extraction requires an authorised
audio source and applicable-use review; academic purpose and
deletion alone do not establish permission.
No audio is redistributed.
Temporary analysis audio is
deleted after extraction and quality checks, within a planned
seven-day maximum. Retain only necessary pseudonymous fea-
tures, recording identifiers, checksums, match decisions and ex-
traction/deletion logs under the consented retention policy. Ex-
perimental playback has separate source and retention require-
ments.
7.6 Readiness
The history layer is proposed. Record import, matching, per-
mitted audio access, extraction, behavioral calibration and its
separate precision analysis remain to be completed. The app il-
lustrates the design and collects no history or responses. Ethics
approval and registration precede recruitment; neither memory
diagnosis nor autobiographical inference is claimed.
Additional history safeguards
Track-level listening duration does not identify which moments
were heard.
Prespecify aggregation, recency weighting and
missing-data rules in development data, and report matched-
track and listening-time coverage where possible. Recognition
foils permit false-alarm estimation. Record exports have incom-
plete coverage across devices, services and everyday listening;
structural familiarity in new music is a prediction requiring in-
dependent behavioral validation.
Temporary analysis audio is deleted within seven days of ac-
quisition, including unresolved or failed items. Platform access
requirements remain applicable. Parsers, recording matching,
the profile model and research data collection remain to be im-
plemented and validated.
References
[1] Cheung et al. (2019). Uncertainty and surprise jointly predict musi-
cal pleasure. Current Biology. 10.1016/j.cub.2019.09.067.
[2] Mas-Herrero & Marco-Pallarés (2025). Predictive processes shape
individual musical preferences. PNAS. 10.1073/pnas.2500494122.
[3] Kathios et al. (2024; online 2023). Generating new musical prefer-
ences from multilevel mapping of predictions to reward. Psychologi-
cal Science. 10.1177/09567976231214185.
[4] Müllensiefen et al. (2014). The musicality of non-musicians. PLOS
ONE. 10.1371/journal.pone.0089642.
[5] Mas-Herrero et al. (2013). Individual differences in music reward
experiences. 10.1525/mp.2013.31.2.118.
[6] Woods et al. (2017). Headphone screening to facilitate web-based
auditory experiments. 10.3758/s13414-017-1361-2.
[7] Albury et al. (2026). Stimulus predictability and liking en-
hance auditory–motor encoding and memory for melodies.
10.1111/nyas.70310.
Model provenance
MI Systems Neuroscience manuscript, 22 September 2026; static source
audit, 28 September 2026. The original map and channels are unchanged.
The v3 listening-history specification distinguishes proposed additions
from implemented audio routes. No engine or listener-profile validation is
claimed.
Protocol and materials
Source proposal: v3.0, 7 October 2026. mymusicalmind.io/learning/.
This edition includes the substantive listening-history companion, pre-
serves the four numbered manuscript figures as original vector PDF ob-
jects and leaves the existing figure atlas external. The original six-channel
and 20-Hz specifications are retained; this formatting edition does not
amend the study protocol.
History-data documentation
The source companion records a review on 7 October 2026 of Spotify
export fields, Spotify Developer Policy, Apple Music API scope and
YouTube terms. These links document source-access requirements; no
account connection or audio acquisition is part of this edition.
Learning and the pleasure of identical music | Research proposal · 12 October 2026
4
