Learning Transposition-Invariant Interval Features from Symbolic Music and Audio

  • 2018-06-21 13:35:44
  • Stefan Lattner, Maarten Grachten, Gerhard Widmer
  • 10

Abstract

Many music theoretical constructs (such as scale types, modes, cadences, andchord types) are defined in terms of pitch intervals---relative distancesbetween pitches. Therefore, when computer models are employed in music tasks,it can be useful to operate on interval representations rather than on the rawmusical surface. Moreover, interval representations aretransposition-invariant, valuable for tasks like audio alignment, cover songdetection and music structure analysis. We employ a gated autoencoder to learnfixed-length, invertible and transposition-invariant interval representationsfrom polyphonic music in the symbolic domain and in audio. An unsupervisedtraining method is proposed yielding an organization of intervals in therepresentation space which is musically plausible. Based on therepresentations, a transposition-invariant self-similarity matrix isconstructed and used to determine repeated sections in symbolic music and inaudio, yielding competitive results in the MIREX task "Discovery of RepeatedThemes and Sections".

 

Quick Read (beta)

loading the full paper ...