Can We Trust What We Hear? Advancements in Deepfake Audio Detection

Recent advances in generative AI, particularly in speech synthesis and voice conversion, have enabled the creation of highly realistic deepfake speech, posing significant risks to security, privacy, and trust. Building on recent progress in this area, this talk presents an overview of state-of-the-art techniques in deepfake speech detection, with emphasis on both algorithmic innovations and practical applications. We will examine modern approaches based on deep neural architectures, self-supervised learning, and acoustic feature analysis, along with key challenges in maintaining robustness against the evolving nature of speech deepfakes. Beyond speech, the talk will briefly extend to emerging research in deepfake environmental audio, highlighting how detection strategies can be adapted to broader audio domains. Finally, the presentation will outline open challenges and future directions toward developing trustworthy and deployable audio deepfake detection systems.