Overview
ASVspoof studies how to distinguish genuine speech from replayed, synthesized, converted, cloned, or partly manipulated audio, helping prevent forged voices from fooling speaker-verification systems. Methods range from handcrafted spectral cues and RawBoost-style augmentation to self-supervised encoders such as WavLM and XLS-R, paired with CNN, Transformer, or graph back ends, temporal localization, and score fusion. Evaluation uses logical-access, physical-access, and deepfake tracks in ASVspoof benchmarks, alongside cross-dataset and in-the-wild tests, with metrics such as EER and minDCF supplemented by calibration and robustness analyses. Strong benchmark results remain hard to transfer across codecs, channels, unseen generators, and demographic or prevalence shifts, leaving open how to improve generalization and auditable decisions without relying on dataset shortcuts or costly ensembles.
Latest papers
31 papers6 papers in the last four weeks, up 20% on the four weeks before. 0.0% of all new papers.