Skip to content

GuitarSet

Summary

~3 hours of solo acoustic guitar recordings with comprehensive annotations: note-level (pitch, onset, offset), string-level, fret-level, and playing technique annotations. Recorded with hexaphonic pickup providing per-string audio alongside mono microphone. 360 excerpts across 6 players and multiple styles. JAMS annotation format. Presented at ISMIR 2018 (150+ citations). GitHub: marl/GuitarSet (160 stars).

Key Claims

  • Hexaphonic pickup provides ground-truth per-string audio for training separation/transcription models
  • JAMS format standardizes multi-level annotations (notes, strings, frets, techniques)
  • 3 hours sufficient for training initial deep learning models (as demonstrated by TabCNN)
  • Dataset enables both transcription and source separation research on guitar

Relevance to Banjo

The blueprint for creating "BanjoSet." Every aspect of the GuitarSet methodology translates to banjo: - Per-string pickup (piezo or magnetic split pickup on banjo) provides ground-truth string-level audio - JAMS annotation format works for any stringed instrument - Same annotation schema: pitch, onset, offset, string, fret, technique - ~3 hours likely sufficient for initial banjo transcription model training

A banjo dataset should additionally capture: drone string activity (5th string, almost always open), roll patterns, and bluegrass-specific techniques (hammer-on, pull-off, slide, choke).