Reproduce “Elucidate Gender Fairness in Singing Voice Transcription” 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

Xiangming Gu, Wei Zeng, Ye Wang

Researchers from the National University of Singapore empirically demonstrated that state-of-the-art singing voice transcription (SVT) systems consistently exhibit a performance bias favoring female voices, with pitch differences as the primary cause, leading to F1-score gaps of 7-15%. They introduced Note-Conditioned Adversarial Learning, which reduced this gender bias by over 50% across multiple datasets while incurring less than 2% degradation in overall transcription utility.

Key insights

  • Current state-of-the-art SVT systems consistently show superior performance for female singing voices, with significant pitch-related F1-score disparities (7-15% gaps) favoring females.
  • Gender bias in SVT is primarily caused by differences in pitch distributions between male and female voices, rather than merely data imbalance in training sets.
  • Conditioning bias mitigation on task-relevant features like note events, as in NCAL, is effective for reducing bias while maintaining high utility in SVT systems.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.