Cover Song Retrieval / Build notes
Version identification, written from the code that runs.
Posts from a cover-song retrieval pipeline on Da-TACOS: what chroma keeps and what it throws away, how key and tempo invariance are actually handled, and why the evaluation protocol decides more of the result than the model does.
Upcoming
- Why chroma is the right representation for cover detection, and what it costs you.
- The Optimal Transposition Index: a twelve-line idea that does most of the key-invariance work.
- Subsequence DTW from scratch, and the length normalisation nobody mentions.
- Writing MAP, MRR, and Recall@K by hand, and testing them against a ranking worked out on paper.
- WID versus PID: how a split at the wrong granularity quietly invalidates a benchmark.