Chapter 6 · measured
Four detectors, four cells, three metrics
A cell is one experimental condition. Reading across a cell is the accent change; reading down it is the codec. The contrast that matters is the difference of the two differences — the interaction.
Cell means
Every detector loses more to the codec than to the accent
Bars are the seed mean; whiskers are one standard deviation. AASIST ran a single seed, so it has no whisker — its flat profile is a failed instrument, not a stable detector.
| Detector | seeds | A | B | C | D | Δacc | Δch | Δint | 95 % interval on Δint | verdict |
|---|
All values are percent EER. Δacc = C − A, Δch = B − A, Δint = (D − B) − (C − A). The interval is the envelope of the per-seed speaker-cluster bootstrap intervals, not a pooled interval over seeds.
The three contrasts
What the accent costs depends on how much the codec already took
If the two harms were independent, Δint would sit at zero. It does not: for MelCNN the accent change is cheaper after compression than before it, because the compressed cells have less headroom to lose.
Secondary gate
Does accent-aware training actually help on speakers it has never seen?
Three arms on the same held-out set of German books: the plain detector, a size-matched control that sees the same amount of data without the accent shift, and the accent-trained arm. The gate was pre-registered at three EER points; every seed clears it.
| Arm | seed 0 | seed 1 | seed 2 | mean | vs plain |
|---|
Cell C EER (%), held-out German books with the shifted accent. The size-matched control moves the mean by a fraction of a point and the accent arm moves it by several, so the gain is the accent data and not the extra utterances.
Provenance
Every run, including the ones that are read differently
Nothing was discarded. A row that is not in the chapter is still a row that was run; the corpus tag says how each is read.
| Corpus | Detector | seed | A | B | C | D | Δacc | Δch | Δint | verdict |
|---|