Conversation

Notices

1. 🇳🇴 Thor — backup account (thorthenorseman@octodon.social)'s status on Sunday, 19-Aug-2018 04:08:51 EDT 🇳🇴 Thor — backup account

   I'm pretty sure I could code a music recommendation algorithm that actually works. It would pretty much ignore all the metadata and analyse the audio waveforms instead, and pick up on things like timbre, energy, noise/sizzle, tempo, rhythm, instruments and vocals, and create a multidimensional vector for every song in the library, and then I'd match that against one or more equivalent vectors belonging to the current user in order to make recommendations.

   1. 🇳🇴 Thor — backup account (thorthenorseman@octodon.social)'s status on Sunday, 19-Aug-2018 04:11:53 EDT 🇳🇴 Thor — backup account

      in reply to

      There are sweet spots for all of those parameters that I think describe a person's music tastes pretty well. The only thing it couldn't really help you with is good lyrics, for those that care about lyrics. I suppose you *could* use metadata for that and use a Bayesian classifier to give every lyric one or more personalised probabilities that you include in the multidimensional vector.