Reproduce “MERIT: Learning Disentangled Music Representations for Audio Similarity” 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
Abhinaba Roy, Junyi Liang, Dorien Herremans
MERIT gives music-search systems separate similarity scores for melody, rhythm, and timbre, illustrated by a piano cover that preserves a rock song’s melody while changing its arrangement. It uses three trained projection heads on a shared frozen MERT encoder. On held-out factor-controlled triplets, the rhythm head reaches 100.0% rhythm accuracy, but this result does not establish selectivity on unrestricted real-world music.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.