Conversation

Notices

1. Embed this notice
   Rasmus Fleischer (rasmusfleischer@tldr.nettime.org)'s status on Sunday, 05-May-2024 21:40:50 JST Rasmus Fleischer

   Wikipedia on #Udio:
   "Critics praised its ability to create realistic-sounding vocals while others raised concerns over the possibility that its training data contained copyrighted music."
   https://en.wikipedia.org/wiki/Udio

   What I have not yet seen discussed is the issue of data annotation. Recorded music is worthless as training data if it is not connected to words describing it.
   Are there any clues as to how companies like Udio gather annotations that describe pieces of music?
   To what extent can be say that a service like Udio exploits the labour not only of musicians, but also music critics?
   Have annotation data been extracted from playlist names on platforms like Spotify and Youtube, tags on Soundcloud or Bandcamp, etc? And to what degree may the annotations themselves be AI generated?

   I have a vague feeling that there is so much to unwind here.

   • Embed this notice
     Rasmus Fleischer (rasmusfleischer@tldr.nettime.org)'s status on Monday, 06-May-2024 16:25:03 JST Rasmus Fleischer

     in reply to

     • Jonas Bjärnstedt

     @jbjarnstedt Thinking of any particular projects?

   • Embed this notice
     Jonas Bjärnstedt (jbjarnstedt@eupolicy.social)'s status on Monday, 06-May-2024 16:25:05 JST Jonas Bjärnstedt

     in reply to

     @rasmusfleischer All the tagging projects that I've seen in recent years, in different sectors, suddenly seems less benign.