Silence, Channels and Calibration: A Dual-Model WavLM System for Synthetic and Replayed Speech Detection

Voice deepfakes and replayed recordings threaten speaker verification, banking and telephone-based authentication, yet detectors that score well on public benchmarks often fail on real recordings. We present a dual-model countermeasure built on a self-supervised WavLM Base+ encoder with a learnable weighted sum over all 13 hidden layers, attentive statistics pooling and a One-Class Softmax objective: one model targets synthetic speech (ASVspoof 2019 LA) and a second targets loudspeaker replay (ASVspoof 2019 PA). We show that the two tasks require opposite augmentation policies, since channel distortions are nuisance variables for synthetic-speech detection but the very evidence of replay. On the official LA evaluation protocol our synthetic-speech model reaches an equal error rate (EER) of 3.36% and a minimum tandem detection cost function of 0.0853, outperforming both challenge baselines by more than a factor of two. A controlled 2 × 2 experiment reveals that this score depends heavily on the known silence artefact of the corpus: trimming leading and trailing silence raises the evaluation EER to 16.32% and flags 31.5% of genuine speech as fake. Retraining on trimmed audio removes the dependence entirely (17.38% vs. 17.51% EER with and without silence) and exposes the true difficulty of unseen neural attacks. On the real-world In-the-Wild corpus the system obtains 19.31% EER, well below the 33.1% previously reported for an ASVspoof-trained RawNet2. The replay model generalises from simulated to real replays with only a 3-point EER loss (23.32% to 26.35%), and its errors are explained almost entirely by loudspeaker quality. A logistic calibration on an atanh-stretched score reduces calibration error and balances operating-point errors. The complete system is deployed as an explainable, offline-capable dashboard.