started thinking through a recommender that takes a song, a snippet, a music file, or even a youtube link with a timestamp in the middle of a guitar strum, and finds other music that hits the same nerve. not mainstream discovery, the niche stuff that lives in corners of bandcamp and instagram reels. my first instinct was the shazam style fingerprint — hash every song, compare distances, the minimum distance is the most similar track. sounds perfect until you learn the details.
a chromaprint is built to avalanche. small acoustic differences flip the whole hash, so two songs that sound alike share basically zero hashes and the distance between their fingerprints is pure noise. it returns random tracks with full confidence. the graded comparison actually lives one step before the hash, in chroma, the raw tonal content, and the right shape for hashing similar things together is an embedding plus a vector index. there's also audiomuse, a self-hosted sonic similarity engine that plugs straight into the music server already running here.
the other hard problem is unreleased music. artists post tracks only on social platforms to build a following, and no recognition database has them. so identification becomes its own resolver layer that runs on every clip: the platform credit first, then the context around the post, then a fallback chain of recognition services with different underground coverage. the artist stays findable even when the track stays unknown, which is the whole point. a vibe lives in the artist as much as in the song.