Abstract

Audio Language Models (ALMs) are evolving, and with their advancements, new challenges in detecting deepfake audio arise. This paper introduces Codecfake, a dataset designed specifically for training and evaluating deepfake audio detection models.

Key Findings:

• The dataset includes diverse examples across two languages to train robust detection models.
• The introduction of domain adaptive techniques to manage the discrepancy in training data for better generality across unseen data.

Implications:

• The development of more robust detection systems could help in early detection and prevention of audio-based misinformation.
• Training on such specifically crafted datasets could set new standards for benchmarking deepfake detection models.

Opinion: This study underscores the critical need for focused efforts on developing tailored datasets for emerging threats like deepfake audio. The methods and dataset presented could serve as a blueprint for future research in multimedia security.