AUDITORY LEARNING · SPEECH · PROSPECTIVE RESEARCH PROTOCOL
Can learned voice cues improve speech recognition in noise?
Amaç Erdem | amace@bu.edu | ORCID 0009-0000-1711-0036
Abstract. Familiar voices can improve speech recognition in noise, but this benefit does not identify which learned acoustic features are useful
or how to train them. We propose a speech study in adults with sensorineural hearing loss. A measured bank of pre-generated utterances will
support contrasts in spectral structure and segmental timing, with fundamental frequency constrained. Randomized familiarization, reciprocal
component comparisons and separate-day replication will test learning-dependent cue use. A subsequent trial will compare profile-guided
training with equally intensive control training. Its primary endpoint is recognition of new natural talkers four weeks after training. Clinical
efficacy and the incremental predictive value of Musical Intelligence will be evaluated separately.
Academic edition r3 · 12 October 2026 · Source: draft 1.1, 5 October 2026. Unregistered proposal; no participant results.
Spectrotemporal profiling and individualized auditory training
Keywords: speech-in-noise; talker learning; spectral envelope; temporal cue use; auditory rehabilitation
1. Question and prior evidence
Which spectral and temporal features of a learned talker improve
recognition when carried into new speech? Do their contribu-
tions combine additively or interact? Can that profile identify
training targets that produce retained gains for previously un-
heard talkers?
Familiar-talker benefits have been demonstrated in older lis-
teners with hearing loss [1]. Laboratory familiarization can also
improve intelligibility and reduce listening effort [2].
These
findings motivate learning as an intervention target; they do not
establish which acoustic components produce the benefit.
Speech-cue weighting varies across listeners with hearing
loss [3], and changing masker modulation shifts sentence-level
spectral weights [4]. A listener’s profile is therefore conditional
on the task and acoustic scene. The objective is better recogni-
tion, not equal spectral and temporal weights.
2. From music to speech
Originating in musical source selection, the proposal asks
whether learned acoustic structure helps listeners follow a tar-
get amid competition. Speech makes that benefit directly mea-
surable as word recognition; transfer from music training is not
assumed.
3. Testable hypotheses
H1 · Learned component use. Familiarization changes the con-
tribution of at least one prespecified spectral or timing contrast
beyond matched exposure to a control voice. Both adding a B-
like component to A-like speech and replacing it in B-like speech
are tested.
H2 · Joint dependence. Familiarization changes the spec-
tral × timing interaction. Its sign is not prescribed; a null result
is informative only if uncertainty excludes a prespecified mean-
ingful interaction.
H3 · Retained transfer. A frozen profile-guided selection
policy improves speech reception threshold for new natural talk-
ers four weeks after training, relative to a dose-matched active
policy.
4. Research approach
Spectral-weighting methods provide a direct measurement
precedent [4]. A rhythm/EEG preprint links target/background
timing to recognition and neural tracking [5].
The proposed
bridge is to ask whether learned component use predicts train-
ability and retained transfer.
The project combines auditory assessment, learning experi-
ments and computational modelling to relate individual cue use
to speech recognition.
Table 1. Established starting points and the proposed next test.
Evidence
What it supports
What remains to be tested
Familiar-talker benefit [1, 2]
Recognition depends on experience
Which learned components contribute?
Context-dependent cue weights [3, 4]
Cue use varies with listener and scene
Is the learning-related profile repeatable?
Targeted auditory practice [6]
Task-level gains are possible
Does profile-guided practice transfer?
5. MI resolves features and qualifies the stimulus con-
trasts
ElevenLabs will generate matched-text candidates. R3/T3 will
measure the resulting audio, preserving the complete feature
vectors and extraction settings. Two recordings with the same
plotted distance may differ in their component measurements;
X/Y proximity is therefore insufficient for experimental accep-
tance.
The stimulus pilot will select a small set of spectral and tim-
ing contrasts. Each comparison must achieve its intended feature
change while bounding off-target changes. The proposed four-
cell design proceeds only if enough recordings satisfy these rules
across sentences and talker pairs.
5.1 Specify features before testing recognition
R3 contributes envelope, contrast, flatness, flux and modulation
channels. These are computational descriptors, not direct for-
mant or phoneme measurements. A proposed speech layer adds
validated F0, formant, aperiodic-energy and alignment measure-
ments. Preserve durations on the original clock when aligning
spectral comparisons. Preregister the target coordinates and nui-
sance set; if several terms remain coupled, interpret their joint
contrast.
5.2 Qualify comparisons, not just recordings
Retain per-feature target deviations, nuisance deviations and
missing-value flags relative to both references. Fix scaling and
acceptance tolerances in the independent pilot; audit residual co-
variance and all four cells across texts. Recheck after playback
or hearing-aid processing [7]. Failed candidates return to gen-
eration or resynthesis. Passing the audit supports a specified
acoustic comparison; it does not establish perceptual indepen-
dence.
6. Separate component learning from clinical transfer
Stage I will recruit English-speaking adults with bilateral sen-
sorineural hearing loss and an age-matched normal-hearing
comparison group. Audiometry, speech-in-noise performance,
device use and language background will characterize the sam-
ple. Cochlear implant users will not be pooled with acoustic-
hearing participants.
Talkers are counterbalanced across roles. Participants are
randomized to familiarization with B or matched exposure to a
third talker C, absent from the A/B test bank. A is an untrained
reference, not a zero-feature voice. New sentences are used in
each pre-training, post-training and replication session.
Learned voice cues and speech recognition | Research proposal · 12 October 2026
1


From waveform to qualified contrasts
MI ENGINE / FEATURE-RESOLVED MEASUREMENT AND STIMULUS ACCEPTANCE
a
One signal, a traceable measurement record
/pɑ/
/si/
/ku/
/tɛ/
/ʃa/
/mi/
Matched sentence · references A / B
Audio hash + voice / generator version
Fixed extraction settings + alignment
b
R³ · spectral components
c
T³ · temporal organization
Harmonics (a.u.)
Aperiodic (a.u.)
0
2
4
Frequency (kHz)
0
3
6
Frequency (kHz)
Reference
Changed spectrum
F0 = 145 Hz; amplitudes redistribute.
MFCC [63–75] · contrast [76–82] · flatness [23]
Flux [21] · onset [11] · modulation [83–88]
Selected feature trajectory
rf(t)
200 ms
600 ms
2,000 ms
−2.0
−1.0
0
Time relative to t (s)
Mean · variability · change; L0 = past-directed.
d
Speech-specific integration
PROPOSED ADAPTER
Checked phoneme timing · F0 · F1–F3 · harmonic / aperiodic measures · calibrated level
e
Qualify the contrast
f
Require all four cells
Target terms
Other terms
Intended displacement
Within fixed tolerances
S comparison: bound timing. T comparison: bound spectrum.
Both: F0, level, pronunciation, naturalness and covariance.
Sᴬ Tᴬ
Sᴬ Tᴮ
Sᴮ Tᴬ
Sᴮ Tᴮ
FAIL → regenerate / resynthesize → re-measure
PASS → full vectors for analysis; X/Y for display
Procedural illustrations; no measured spectra, accepted research cells or participant results.
Fig. 1. MI as the measurement and stimulus-audit layer. a–c, A schematic utterance, harmonic/aperiodic spectra and proportional window supports explain existing
R3/T3 operators (zero-based indices). d–f, Proposed speech measurements feed target/off-target acceptance and four-cell qualification. The intervals express rules,
not observed measurements. Analysis retains full vectors; X/Y provides a display. Stimulus selectivity and coverage remain untested.
6.1 Mapping in competing sound
Primary mapping uses two competing talkers. A separately anal-
ysed speech-shaped-noise condition tests masker dependence.
Set an individual challenge level before familiarization, then
hold it fixed across that listener’s matched cells. Target cues
contain different words from the scored sentence. Counterbal-
ance masker identities and trial order.
The listener types the target words.
Primary scoring is
open-set word accuracy; intrusions, voice identification and re-
ported effort are secondary. Store raw responses and prespecify
spelling normalization. Word errors are not automatically at-
tributed to a phoneme or cue.
6.2 From a profile to a training target
Re-estimate spectral, temporal and joint effects on a separate
day and new utterances. Then use independent, counterbalanced
learning probes to test whether each candidate contribution can
change. A small baseline effect is insufficient grounds for tar-
geting it.
The policy uses repeatability, uncertainty and probe re-
sponse. Thresholds and eligibility rules are fixed in Stage I; re-
port the eligible proportion. Any benefit applies to this defined
population of eligible profiles.
6.3 Independent policy trial
In a new SNHL cohort, eligible participants are randomized 1:1
to profile-guided or balanced control selection. Both arms re-
ceive the same task families, talker pool, feedback, duration and
masker progression. The intended dose is 20 minutes/day, five
days/week for four weeks, subject to feasibility confirmation.
Begin near the learned reference and introduce accepted
recordings with reduced familiar support. Guidance may priori-
tize spectral, temporal or joint variation according to the locked
rule. The control uses a prespecified balanced schedule; the ma-
nipulated factor is selection policy.
6.4 Outcome and retention
The primary endpoint is baseline-adjusted speech reception
threshold (SRT) at week 8, four weeks after training ends. Test
natural talkers and sentences never used in profiling or training;
lower SRT is better. Assessors are blinded. Immediate post-
training performance is secondary.
The goal is retained speech recognition in noise.
Better
recognition does not demonstrate recovery of cochlear sensitiv-
ity or normalization of a clinical audiogram.
Learned voice cues and speech recognition | Research proposal · 12 October 2026
2


Explore the stimulus space
INTERACTIVE 3D COMPANION · Original static figure below.
a
A discrete, reference-relative bank
b
Speech structure at fixed F0
F
Anchors
Z · competing sound
TMR: +12 → −12 dB
X · spectral distance
Y · temporal distance
F0 eligibility → select stored recordings
X/Y: normalized distances. Gray cloud: illustrative masker.
0
2
4
6
kHz
0
1
2
3
Time (s)
Harmonic
Aperiodic
Formants
F0 = 145 Hz · harmonic spacing preserved
Procedural structure, not an STFT or measured MI output.
c
Timing redistributes segments and pauses
Reference
Altered timing
/pɑ/
/si/
/ku/
/tɛ/
/ʃa/
/mi/
Same event order and total span; local timing changes.
0–3 s
Interactive companion. The stimulus map and its acoustic interpretation. a, X/Y select an F0-compatible bank anchor; Z specifies target-to-masker ratio after
target selection. b, The procedural speech model shows harmonics, formant emphasis and aperiodic consonant energy at fixed F0. c, Timing redistributes segment
durations and pauses at fixed utterance span. All points, curves and amplitudes are schematic. The companion web article animates the same controls; it contains
no generated research audio or MI measurements.
Open the interactive 3D figure: mymusicalmind.io/speech#space. Drag to rotate and move X, Y and Z in the web view. The PDF preserves the original static
companion.
7. Learning-specific contrasts
Fit word-correct responses with a hierarchical logistic model
including component cell, exposure arm, session and hearing
group. Model listener, sentence, talker-pair and masker varia-
tion. Let qst denote estimated log odds for spectral state s and
timing state t:
𝐶𝑆= [(𝑞10 −𝑞00) + (𝑞11 −𝑞01)]/2,
𝐶𝑇= [(𝑞01 −𝑞00) + (𝑞11 −𝑞10)]/2,
𝐼= 𝑞11 −𝑞10 −𝑞01 + 𝑞00.
The learning-specific effect is the pre-to-post change in CS,
CT or I after B training minus that after C training.
Report
probability-scale marginal effects as well. Retain the prespec-
ified feature vectors in the model; report bundle effects where
terms cannot be separated. These contrasts measure dependence
on the accepted stimulus manipulation; they are not percentages
of attention.
8. Clinical and computational tests
For H3, estimate the baseline-adjusted between-arm SRT differ-
ence at retention, using intention-to-treat analysis and prespec-
ified missing-data sensitivity analyses. Improved trained-item
performance alone is insufficient.
For MI, add T3 summaries to a matched acoustic/exposure
baseline.
Candidate native spans of 200, 600 and 2,000 ms
are modelling choices, not attention categories.
L0 is past-
directed; centred R3 frames and clip normalization preclude on-
line causality. Tune in development folds; test log loss and cali-
bration on held-out listeners and talkers.
9. Optional neural extension
After behavioral feasibility, separate continuous-speech blocks
may test target-versus-masker EEG tracking under acoustically
counterbalanced conditions. The rhythm and neural-tracking
preprint supplies a relevant precedent [5], not proof that im-
proved tracking mediates rehabilitation.
EEG remains sec-
ondary to recognition and transfer.
Available fNIRS could address slower prefrontal effort-
related responses, with short-separation and physiological con-
trols [8]. It cannot identify a 150-ms attention window. Neither
modality is needed for the primary policy trial.
10. Decisions before registration
An independent pilot will establish stimulus quality, profile pre-
cision and task burden. Simulation-based sample sizing must
cover both the component interactions and the retained between-
arm SRT contrast. Fix meaningful-effect bounds, stopping rules,
exclusions, multiplicity control and missing-data handling be-
fore confirmatory enrollment.
Obtain ethics approval, voice-use permission and licensed
assessment materials. The bank generator, feature definitions,
model versions and selection policy must be versioned. The
present interface is a demonstrator: no participant, generation
or MI-extraction results are reported.
Learned voice cues and speech recognition | Research proposal · 12 October 2026
3


a
Reciprocal component comparisons
b
Independent transfer trial
A-like timing
B-like timing
A-like
spectrum
B-like
spectrum
00
Reference
01
Timing
10
Spectrum
11
Joint
B familiarization vs control-voice C exposure
Pre → post → separate-day replication
New utterances · accepted recordings · schematic glyphs
STAGE I
Profile → probe → lock
New eligible SNHL cohort · randomize 1:1
Week 0
Week 4
Week 8
Profile-guided
Dose-matched control
Training
No training
PRIMARY ENDPOINT
New natural talkers · speech reception threshold
Four weeks after training ends
No assumed benefit; test the between-policy difference.
Fig. 2. Learning and transfer require different controls. a, Spectral stems and timing strips depict the intended A-like/B-like component combinations; the glyphs are
schematic. Reciprocal contrasts add or replace learned features in accepted recordings. b, Replicated, modifiable profiles determine a frozen selection policy. Its
efficacy is tested in an independent cohort against active training, with the primary retention endpoint four weeks after training ends.
Table 2. Findings that distinguish the proposed claims.
Observed pattern
Permitted interpretation
Component learning; no retained transfer
Learning is task-specific; clinical hypothesis unsupported
Retained transfer; no predicted cue change
Policy may help; the proposed component account is weakened
Behavioral benefit; no MI predictive increment
Clinical result stands; incremental model claim fails
Confounded cells or imprecise effects
Mechanistic test is inconclusive; redesign or more precision is needed
References
[1] Souza P et al. (2013). The advantage of knowing the talker. J Am
Acad Audiol. 10.3766/jaaa.24.8.6
[2] Baxter F, Smith HJ, Holmes E (2025). Voice familiarization training
improves speech intelligibility and reduces listening effort. Trends
Hear. 10.1177/23312165251401318
[3] Souza P, Gallun F, Wright R (2020). Contributions to speech-
cue weighting in older adults with impaired hearing. JSLHR.
10.1044/2019_JSLHR-19-00176
[4] Shen Y, Langley L (2023). Spectral weighting for sentence recogni-
tion in steady-state and amplitude-modulated noise. JASA Express
Lett. 10.1121/10.0017934
[5] Schell E, Cheng TZ, Shen Y, Zhao TC (2026). Rhythm modulates
perception and neural tracking of speech in a speech-in-noise task.
PsyArXiv preprint, v2. 10.31234/osf.io/h7duq_v2
[6] Koprowska A et al. (2023). The effect of phoneme-based auditory
training on speech intelligibility in hearing-aid users. Int J Audiol.
10.1080/14992027.2022.2135032
[7] Yun D, Lentz JJ, Shen Y (2024). The noise reduction algorithm may
not compensate for the degradation in output SNR caused by wide
dynamic range compression. Am J Audiol. 10.1044/2024_AJA-24-
00011
[8] Yücel MA et al. (2021). Best practices for fNIRS publications. Neu-
rophotonics. 10.1117/1.NPh.8.1.012101
Protocol and materials
Source proposal: draft 1.1, 5 October 2026. mymusicalmind.io/speech.
This corrected edition restores the original stimulus-space companion
alongside the two numbered manuscript figures. All three are preserved as
original vector PDF objects with their original content and captions. The
live 3D workspace is linked from the companion in the main text; the PDF
contains its original static illustration.
Learned voice cues and speech recognition | Research proposal · 12 October 2026
4
