three (related) new features just dropped on plyr.fm!
recommended tracks for playlists
recommended tags for tracks
semantic search for tracks (experimental)
all fundamentally powered by the same thing, audio embeddings!
read more about how this works in the docs, but essentially i am running CLAP on modal with turbopuffer to store embeddings to power recommended tracks and semantic search
and i am using effnet-discogs to do recommended tags for tracks.
recommended tracks for playlists
when editing a playlist, a set of recommended tracks will appear based on the tracks in the existing playlist
for example, take this playlist that contains spoken word prayer, entirely in Portuguese:
since all Portuguese prayer on plyr.fm is in this playlist, the most similar tracks are other (Public Domain) audiobooks excerpts that've been uploaded by - the heaviest user of plyr.fm :)
the current implementation works a bit differently depending on the number of tracks in the playlist to extend:
just one: just do top_k where
k=3on the one track2-5 tracks: top_k on each track and RRF on the result
6+ tracks: K-means into <3 centroids, then RRF on centroids
here's a video demo!
recommended tags on tracks
at track upload time, you can select an option to accept auto-suggested tags
or instead in the track edit modal, you are presented relevant tags:
tracks are processed by effnet-discogs in the background on upload or on-demand via GET /tracks/{id}/recommended-tags. predictions (genre tag + confidence score) are cached keyed on audio content because someday i'm going to want to add support for swapping out the audio file for a given track you've already uploaded.
another video demo!
semantic search for tracks (experimental)
hit 🔎 or Cmd + K to try it out
results here are not as good as i want yet. still working on this
as such, this is behind a per-user feature flag. if you'd like to try it and offer feedback, please lmk and i'm happy to open it up to you.
listen to the changelog!
mood search and auto-suggesting tags (and more)
recommended tracks for playlists (and more)
if you've any questions or concerns, please hmu at !