grande verse / taste
tasteyour ariana grande listening profile, and what to hear next
keep a few songs and get something back that is actually about you: which eras your ear leans toward, and ten songs by other artists you probably have not heard, each with a stated reason.
loading the model…
how it works
the hard part of a recommender is not the similarity search, it is having something to compare against. every candidate here comes from acousticbrainz, an open, cc0 database of acoustic analyses, joined to musicbrainz for the metadata. no audio is downloaded, decoded or stored by this site — the descriptors were computed by that project years ago and are published openly, which is the whole reason this page can exist without touching a single copyrighted recording.
those descriptors are fed through a small neural network trained for one job: put songs that belong together close to each other. it is three layers wide and outputs sixty-four numbers per track, which is small enough that the entire catalogue fits in a file about the size of a photograph. the browser downloads it once and does the rest.
the ranking is not simply the ten nearest tracks. nearest-neighbour lists have a well-known failure: the closest songs to any point tend to be by the same artist, so you ask for something like into you and get nine more singles by whoever happened to be nearest. the list is re-ranked so each pick has to be different from the ones above it as well as close to you.
what it does not do
it does not know what you have actually listened to, because it never asks and there is nothing to ask. it does not read lyrics — no page on this site does. and it does not claim the recommendations are objectively correct: they are the closest tracks in a space built from acoustic descriptors, which is a real and checkable claim, and a much narrower one than knowing your taste.
questions people ask
how does it decide what to recommend?
every song in the catalogue was turned into a 64-number vector by a model trained on acoustic descriptors — tempo, timbre, harmony, loudness and how much each of those moves through the track. the songs you keep are averaged into one vector that points at your taste, and the recommendations are the closest tracks to it by cosine similarity, re-ranked so you get ten different songs rather than ten by the same artist.
does it send my choices anywhere?
no. the vector file is downloaded once and every calculation after that happens on your own machine. your keeps and skips are held in this browser's local storage and there is no account, no request and no analytics attached to them.
why does it not recommend her own songs?
because you have just told it which of her songs you like, so handing them back would be the one answer it already knows you have. her catalogue is the question; everything else is the answer.
what does the era percentage actually mean?
each era has a centroid — the average direction of the songs on that record. your taste vector is compared against all eight and the result is turned into a distribution, so you get "41% sweetener" rather than a single label. it describes which records your ear leans toward, not which one you have played most.
why do some recommendations have no reason listed?
a reason is only shown when the recommended track genuinely shares a trait with most of what you kept. if there is no honest overlap the line is left off rather than filled with something that sounds plausible.
further reading
longer pieces on this, written to be checkable — every claim sourced and stamped with the date it was last verified.
- the twenty best songstwenty songs, ranked on the record rather than the receipts. the biggest hit of her career is not number one here, and the reason is in the writing.
- the deep cutspov, ghostin, my hair, honeymoon avenue. fifteen non-singles worth the detour, and what to play first.
- the influencesmariah, whitney, imogen heap, judy garland. who she names, and what is actually audible from each.