Although, would be curious how good you could get to isolating to a single artist. If you had say one exemplar fingerprint per artist, could an out of dataset fingerprint from their discography cluster to that artist? Obviously not for artists who transitioned musical styles.
Or is the algorithm more feature hash than a clusterable feature vector?
Isolating a single artist based on a fingerprint sounds challenging but interesting.
Using exemplar fingerprints, a representative sample of an artist's music, is a good approach, but success would require detailed fingerprints, a varied dataset, and a well-chosen algorithm.
For artists who change styles, time-series analysis can capture their evolving sound.
The solution will likely need machine learning.
The current solution doesn't use feature hashing or clusterable feature vectors. Instead, it relies on audio fingerprinting, which breaks down short audio samples into unique patterns or "fingerprints" for quick comparison with a large database of known songs.
That's like saying the Hutter prize is useless for anyone who doesn't want highly compressed versions of Wikipedia. The underlying code or algorithm is still interesting to study, use, and remix.
AFAIK the underlying algorithm had been implemented multiple times with accompanying explaining blog posts and articles. So this is yet another iteration, done for fun. Nothing wrong with it, just nothing really useful/new I believe.