Speaker: Alejandro Delgado Montero.
Abstract: The recent Environment-Aware Speech and Sound Deepfake Detection (ESDD2) challenge addresses component-level audio deepfake detection across five classes of independent speech and environmental manipulations. We propose systems applying Multi-Head Feature Aggregation (MHFA) to pretrained self-supervised learning (SSL) encoders. We evaluate four architectures: a single encoder, an ensemble sharing a WavLM-Large backbone, an Invariant Domain Feature Extraction (IDFE) architecture, and a logistic regression ensemble of three independent MHFA models. This final ensemble achieves a peak Macro-F1 of 0.6448 on the test set. Class-wise F1 analysis exposes an asymmetry: environmental deepfakes are robustly detected, while concurrent dual manipulations cause acoustic masking, severely degrading accuracy. Diagnostic EER metrics support this trend, confirming that detecting spoofed speech remains the primary bottleneck, while identifying original, unmanipulated audio is highly accurate. Based on: https://www.isca-archive.org/tmp/iberspeech_2026/delgado26_iberspeech.html