I'm trying to find embedding for audio which explains similarities between songs, and have extracted relevant, MFCC, features. From this, I will apply different techniques to find different embedding. My question is as follows: How can one evaluate/explore how "good" an embedding is? What are the different techniques to evaluate/explore? :)

  • $\begingroup$ Do you have any labeled data? Any similarity metrics for pairs, or rankings like 'Song1, Song2 are more similar than Song1, Song3'? $\endgroup$ Sep 19, 2017 at 21:13
  • $\begingroup$ Are you thinking of one embedding (vector representation) per song or is it per word of the lyrics? $\endgroup$
    – ksha
    Sep 20, 2017 at 5:09
  • $\begingroup$ I have rather bad labels in forms of genre. The goal is not to predict genre, but find good embedding for songs. Yes, I'm looking for one vector representation per song. $\endgroup$ Sep 20, 2017 at 14:17


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.