I'm sooo excited! Yesterday, I was fiddling with ways to improve the playlist engine of flowvi.be and had an absolute breakthrough that will cut the server costs by 90%. Not only it makes better playlists but I will be able to offer a one-time lifetime option 💸
This was also quite essential to keep the project sustainable and I want to eventually open source it. I'm also not very keen on making people have yet-another-sub if I can help it.
@iamkonstantin ooh, this sounds great!
I didn’t subscribe when I tried it because I didn’t like the idea of two subscriptions to listen to my music. I’m totally down for one-off payments though!
@iamkonstantin Well done. 👍
@ohjames thank you Mr James
@iamkonstantin Very nice. Can you maybe share what you did? From general perspective?
@CausticHarmony Sure! Some context: Flowvi.be aims to work on device (that's why it's one of very few playlist makers that works offline once it has synced at least once).
There is no AI involved and it needs to work on very small datasets - I never collect or "scrape" user libraries or playlists from the internet. There is a very strong privacy side too - the app doesn't access names of playlists or playlist song information from the user, I don't even "see" which songs are favourited or not.
@iamkonstantin @CausticHarmony If the app doesn’t access playlist song information, how does it know what the song is? Acoustic fingerprinting would take days for my library, so I guess it’s not that.
Or did I misunderstand that part?
@iamkonstantin thank you, very interesting.
@alin
I think this post answers your question:
https://flowvi.be/posts/creating-a-mood-playlist-without-ai/
@CausticHarmony @alin indeed, it's title, artist, album and ISRC (if available) that is read from your personal library during indexing in order to "recognise" which track it is on the server where we associate additional metadata.
The comment about the playlist is that (for example) if you have a playlist called Relax or Workout in Apple Music, flowvi.be won't be peeking into that to improve or learn. It also doesn't look at listening history or any other "personal" markers.
@iamkonstantin @alin very nice, thanks for explaining.
Maybe a longer articlel would be good to answer all those questions?
Also if you need testers for Linux (Mint, Ubuntu) let me know 😀
@iamkonstantin ahh, I see. Reminds me of the days of fixing tags with MusicBrainz and plugins that would add useful acoustic tags like BPM and initial key. Never found a program that would do anything with that information, though. But now we have Flowvibe :)
@CausticHarmony These constraints create interesting technical challenges when generating playlists. When a user connects their Apple Music library, or when they add/remove songs from it, the app needs to "recalibrate" (e.g. recalculate or re-index metadata).
Currently, this happens on the backend and it's a very computationally expensive step. So my research consists in finding ways to bring more of that on device, without consuming too much memory or draining battery.
@CausticHarmony Also, I want to support as many devices as possible and some of them come with very little RAM and CPU. So I can't just "slap a model" on it, it requires engineering and not the kind coding LLMs are good at so it's really quite fun to make progress in that direction :D. Finally, all that needs to be cross-platform enabled, while flowvi.be is currently only released for iOS, a Linux and Windows version are on the roadmap
