Robust AI-Generated Lyrics Detection

The rapid advance of AI-based music generation tools presents new opportunities for the music industry but also poses significant challenges, necessitating reliable methods for detecting AI-generated content. Existing detectors, however, face key practical limitations: audio-based approaches struggle to generalize to unseen generators and are not robust to common audio perturbations, while lyrics-based methods depend on cleanly formatted lyrics that are unavailable in real-world settings. To address this gap, this thesis proposes and evaluates a novel, practically grounded approach that leverages lyrical content extracted directly from the audio signal. Our method first transcribes sung lyrics using a general-purpose automatic speech recognition (ASR) model, allowing established AI-generated text detection methods to be applied. To further improve performance, we introduce DE-detect, a multi-view late-fusion method that also incorporates audio-derived speech features capturing paralinguistic information. By focusing on lyrical and speech-related information rather than low-level audio artifacts, our method is designed for improved robustness and generalization. Experiments on a diverse dataset show that DE-detect achieves strong detection performance compared to text-only ones and, crucially, outperforms audio-based approaches, especially when tested against various audio perturbations and unseen music generators. This work thus presents an effective, robust, and practical solution for detecting AI-generated music.