Wav2Vec Meets Conformer: A Novel Hybrid Approach for Multilingual Deepfake Audio Detection

Deepfake audio refers to synthetic speech that closely mimics a person’s voice, posing risks to security and privacy. This paper proposes a hybrid detection framework combining XLS-R, a multilingual speech representation model, with the Conformer architecture, which captures both local and global audio dependencies. XLS-R extracts rich multilingual embeddings, while the Conformer leverages temporal and contextual features to distinguish genuine from AI-generated speech. Evaluation on benchmark datasets demonstrates that the proposed system achieves improved accuracy and robustness across multiple languages and acoustic conditions.