hey, this is my project, thanks for posting! a little more info:
track similarity is found via cosine similarity between embeddings stored in an approximate nearest neighbor index generated from this model from essentia/upf https://essentia.upf.edu/models.html#discogs-effnet
it's mainly for electronic music discovery and the index includes over a million tracks with that genre sourced via discogs.com data
Hey, just stumbled onto this, it's pretty cool but I figured I'd share my experience on figuring it out
Initial searches:
- "Godspeed": no results for GYBE
- "Pavement": no results for them; seems to be searching by song title
- "Harness Your Hopes": no results... maybe I should try something electronicky
- "Temptation": loads of results but I don't see New Order, I do see Heaven 17 when I scroll way down though so that seems weird
- "Blue Monday": finally!
Once I got a grasp on the vibes it has info on it's pretty cool! Some possible suggestions to build on:
- Having it autoplay onto the next track (would need some logic to skip past broken links too)
- I'd say include something on the landing page about the data that's involved, maybe even with some example acts; when Taylor Swift produces one result and the Beatles none it's probably gonna cause issues
- an option to filter out songs by the same act
- maybe some logic around accented characters; I noticed Beyonce brings up no results but Beyoncé does
Thanks for trying it out and leaving some feedback!
I initially released this to a small Twitter audience who are fans of electronic music and didn't expect it to be posted elsewhere. The plan is to bring in other genres and styles soon while still keeping the data quite niche. My motivation in building this was for music diggers looking for more out-of-the-way suggestions that aren't on traditional DSP catalogues (think sub-genres of dance music that only release 300 copies on vinyl, etc.).
I've updated the h2 tag and some meta descriptions to try flag up electronic only for now and will look at other ways to signpost it better.
The search isn't great, as you mentioned. At the moment, I'm using a simple text matcher from a Vector DB but will look at moving this over to something more powerful that can handle diacritics, ranking, TF-IDF, etc. I have seen a lot of people searching via "artist - track" or pasting a YouTube link for things that are already in the index, indicating they couldn't find it in the search.
track similarity is found via cosine similarity between embeddings stored in an approximate nearest neighbor index generated from this model from essentia/upf https://essentia.upf.edu/models.html#discogs-effnet
it's mainly for electronic music discovery and the index includes over a million tracks with that genre sourced via discogs.com data