Accent robustness and channel compression in audio
deepfake detection — measured separately, then measured together.
Kudomo-Abasi Akpan Ekpo · No. 20254713
Supervisor: Prof. Dr. Fadi Al-Turjman · Nicosia, October 2026
Open with the question, not the method.
Detectors are reported on clean, same-accent speech. Real traffic is neither clean nor same-accent.
The thesis asks whether those two losses add, or overlap.
The problem
Two levers, usually pulled one at a time
The accent belongs to the speaker. The codec belongs to the
channel. A detector validated on one of them has not been validated on
the other, and certainly not on both.
The two harms are not interchangeable.
Accent shift is unavoidable — you cannot transcode a speaker into your training distribution.
Codec is removable if you know the channel — which is exactly why people skip testing it.
The design
Four cells, so the two losses separate
Read across a row for the accent. Read down a column for
the codec. The contrast that matters is the difference of the two
differences:
Δint = (D − B) − (C − A)
If the harms were independent this would be zero.
Say the arithmetic once.
A = matched + clean, B = matched + codec, C = shifted + clean, D = shifted + codec.
Δ int isolates the interaction from both main effects — it is the only number that answers the thesis question.
Result · cell means
Every detector loses more to the codec than to the accent
Seed mean EER by cell, whiskers one standard deviation.
AASIST ran one seed and is shown as a failed instrument.
Let the shape carry it.
B and D are far above A and C for MelCNN and LCNN: the codec dominates.
RawNet2 is the exception — weak everywhere, which is why it is a secondary.
Result · the three contrasts
And the two harms overlap
Δint below zero means the accent change is cheaper after the
codec has already done its damage.
This is the thesis's contribution.
MelCNN: −9.53 points. LCNN: +0.13. RawNet2: +5.16. The sign is detector-dependent.
That dependence is the finding: you cannot state a single interaction number for the field.
Honesty about the interval
−9.53 points, and an interval that admits it is a range
The seed mean is −9.53. The speaker-cluster bootstrap interval is
[−23.86, +6.27]. Two of the three per-seed intervals exclude zero, so
the sign is reliable; the magnitude is not pinned down.
The paper says both, in that order, and does not
let the point estimate stand alone.
Seed mean
−9.53
Interval
[−23.86, +6.27]
Seeds excluding zero
2 of 3
Anticipate the examiner's question.
"Is the effect real?" — the sign is, on two seeds; the magnitude needs more seeds, and that is stated.
Better to be the candidate who names the weakness than the one who is caught by it.
Robustness check
Does the ordering depend on the metric?
AUC agrees with EER in all four cells. Min t-DCF does not:
it inverts the order in cells B and D, and the paper reports that
disagreement rather than choosing the metric that agrees.
Do not oversell agreement.
AUC: monotone, agrees everywhere.
min t-DCF: disagrees in B and D. It is reported as a secondary and never gates.
Intervention
Can you train the accent away?
Arm
seed 0
seed 1
seed 2
mean
vs plain
Accent-aware training recovers
4.51 points on speakers it never saw. A size-matched control moves
the mean by 0.38 — the gain is the accent data, not the volume.
The control arm is what makes this credible.
Pre-registered threshold: 3 points. Every seed clears it.
The size-matched arm rules out "you just trained on more data".
The seminar · systematic review
769 records, 93 included
The flow closes exactly: excluded, carried forward and
included sum to the number screened.
One screener, so no κ is reported. The reliability
figure is a 10 % intra-rater re-screen — 50 of 63 agreements — labelled
as what it is everywhere it appears.
The review's own finding is the exclusion profile.
264 of 370 exclusions failed one criterion: no refusal measurement.
The field names over-refusal far more often than it measures it.
The seminar · why records fell out
Most of the field does not measure refusal at all
Screening exclusions by recorded reason, both databases pooled.
Distinguish "does not measure" from "is bad".
Attack-success papers are good work; they simply do not report a refusal rate.
The review's criterion is measurement, not quality.
Limits
What this does not show
The interaction is detector-dependent. A single number for the
field would be a fiction; the thesis reports four and their sign
disagreement.
Three seeds. Enough to locate a sign, not enough to pin a
magnitude.
One codec family. G.711 is the telephone baseline; wideband
codecs are a different regime.
AASIST could not be evaluated. It failed the instrument check
before the experiment began, and is reported as a failed instrument
rather than a result.
The review has one screener. The calibration sheet is built so
a second screener can be added without re-drawing the sample.
Own the limits before the panel finds them.
Each limitation has a named next step in the completion ledger.
In one sentence
A codec-degraded detector has less to lose from an
accent change — so the two harms overlap, and by an amount that depends
on the detector.
Which means an accent-robustness claim
measured on clean audio does not transfer to compressed audio, and the
direction of the error is not the same for every architecture.
End on the claim, not on the acknowledgements.
Then: thank the supervisor, invite questions.
Thank you
Questions.
Every number in this deck is generated from the repository's result files;
the site and the paper share one source. thesis.html · seminar.html ·
demo.html