When the Channel Changes: Layer Choice, Calibration Transfer and a Padding Pitfall in Frozen-WavLM Speech Deepfake Detection

Speech deepfake detectors are usually reported by their equal error rate (EER) on clean benchmark audio, while the people who would use them hear speech through phone lines, codecs and messaging apps. We train light classifiers on frozen WavLM Base+ features from ASVspoof 2019 LA and study what happens when only the channel changes. Across four hidden layers, two classifiers, three classifier seeds and four conditions (clean, a simulated narrow-band phone line, the same line with 15 dB noise, and Opus at 16 kbps), the choice of layer matters more than the classifier: layer 6 with a small MLP keeps a mean EER of 1.7%, 2.2% and 2.1% on clean, phone and Opus audio, and 8.1% with noise, where layers 3 and 12 reach 23.6% and 27.0%. On the out-of-domain In-the-Wild dataset (31,779 clips) the same model reaches 18.0% EER and keeps the best ranking of layers. Calibration fitted on clean development data does not transfer: Cllr exceeds 1 bit in 5 of 8 WavLM models with noise and in 7 of 8 on In-the-Wild, while minCllr shows that the scores still separate the classes. With the threshold fixed on development data, the dominant error on In-the-Wild is real speech called fake (30.7% of real clips against 7.0% of fakes missed). We also document how zero-padding inside a batch changes WavLM Base+ features through its group-normalised first convolution: for clips shorter than 2 s, the log-likelihood ratio moved by 2.98 on average and 11% of three-way verdicts changed. Code, result files and a browser demo are public.