Same Bits, Different Verdict: The Decoder in Speech Deepfake Detection

Decoding a bit-identical Opus bitstream with libopus 1.6.1’s stronger optional decoder-side post-filter enabled rather than disabled shifts class-mean detector scores toward the synthetic class in 20 of 20 detector–class–corpus cells, with no packet loss and no change to a transmitted byte.

At a threshold transferred from the unenhanced rendering at a 1% false-alarm target, false alarms rise in all eight cells; one public detector goes from 1.00% to 3.80% while its equal-error rate moves only from 0.30% to 0.40%. The shift is not a monotone rescaling that recalibration absorbs: items cross the threshold in both directions, and equal-error rate moves in every cell. On held-out data a stale threshold multiplies false alarms by 1.7 to 3.8.