Reproduce “AI-Generated Song Detection via Lyrics Transcripts” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Markus Frohmann, Elena V. Epure, Gabriel Meseguer-Brocal, Markus Schedl, Romain Hennequin
Researchers from Deezer and Johannes Kepler University Linz developed an end-to-end pipeline to detect AI-generated songs, utilizing lyrics transcribed directly from audio waveforms. This method demonstrates enhanced robustness against audio manipulations and superior generalization to unseen AI audio generators compared to traditional audio-artifact-based detection techniques.
Key insights
- Lyrics, when extracted from audio via ASR, provide significantly more robust and generalizable features for detecting AI-generated content compared to low-level audio artifacts.
- LLM-based text encoders consistently demonstrate superior performance in identifying AI-generated lyrics, even when derived from noisy ASR transcripts, showcasing their ability to capture subtle generative patterns.
- The negligible performance gap between using ground-truth lyrics and ASR-transcribed lyrics highlights the effectiveness of current ASR technology in bridging the modality gap for content detection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.