Skip to content

Guitar Tablature Estimation with a Convolutional Neural Network

Summary

One of the few models that directly outputs tablature (string + fret positions) from audio rather than just MIDI pitches. Uses CQT spectrograms as input through a CNN that outputs a 6x21 (strings x frets) representation. Trained on the GuitarSet dataset. This solves the "fingering ambiguity" problem — the same pitch can be played on different string/fret combinations, and the model learns which is physically likely.

Key Claims

  • Direct tablature output (string + fret) is feasible with a CNN on CQT spectrograms
  • The CQT representation aligns naturally with musical pitch spacing
  • GuitarSet provides sufficient annotation quality for tablature training
  • The model implicitly learns physical fingering constraints from data