MSc thesis · Near East University · 2026
Does an accent survive the codec?
Audio deepfake detectors are usually tested on clean, same-accent speech. Real speech is neither. This study measures the two losses separately — then measures what happens when they land together.
Mean EER over three seeds, MelCNN. Click a cell to see what it costs and why it is there.
The result, in four numbers
Two harms do not add. They overlap.
Every number below is a difference in equal-error-rate points, measured on the paired MLAAD corpus and carried from the same result files the paper's tables are generated from.
Where to go
Three surfaces, one study
The experiment
Four detectors × four cells × three metrics, every run listed, intervals that admit what they do not show. Switch detector, read the contrasts, watch the ordering change.
The systematic review
769 records identified, 571 screened, 93 included. The funnel closes arithmetically; every exclusion carries a recorded reason; the one-screener reliability is stated as an intra-rater figure, never κ.
The deployed service
The detector behind an HTTP endpoint, scored on the same utterances twice: once clean, once through G.711 A-law. What the model does when you actually send it audio.
Presentation mode
A full-screen deck built from the same numbers, driven by arrow keys. Press ? inside it for shortcuts; press P to print it as a handout.
What this site is
A progress report, presented as a progress report. The measurements are finished and reproducible from the repository; the writing is what remains. No number on these pages is typed by hand — the site reads the same results/ files the paper's tables are generated from, and the build script refuses to produce a figure the data does not support.
Where a claim is weaker than it looks, the page says so. The interaction's sign is reliable on two of MelCNN's three seeds, not all three. AASIST failed the instrument check and is reported as a failed instrument, not as a detector result. The accent-aware training gate rests on three seeds, so the interval carries the claim and the mean is shown beside it rather than instead of it.