TabCNN¶
Summary¶
A CNN that estimates guitar tablature (string+fret positions) directly from audio of solo acoustic guitar. This is one of very few models that outputs tablature directly rather than just MIDI pitches — it solves the fingering ambiguity problem (same pitch achievable on different string/fret combinations). Uses CQT spectrograms as input, outputs a 6x21 (strings x frets) representation. Trained on GuitarSet. Presented at ISMIR 2019 (76+ citations).
Key Claims¶
- Direct tablature output (string+fret) is feasible from CQT spectrograms via CNN
- Solves the fingering ambiguity — multiple string/fret combinations for the same pitch
- GuitarSet provides sufficient labeled data for acoustic guitar tab estimation
- CNN architecture is adequate; transformer not required for this task
Relevance to Banjo¶
Most directly relevant architecture for banjo tab transcription. The same approach (CNN on CQT spectrogram -> stringxfret output) can be adapted to banjo by changing output dimensions from 6x21 to 5x22. Banjo is simpler: primarily monophonic, consistent open-G tuning, drone string always open.
Minimal code available — this is architecture inspiration, not a ready-to-use tool. Follow-up work by Cwitkowitz et al. (2022) added string transition feasibility constraints.
Related¶
- ../entities/tabcnn — entity page
- ../entities/guitarset — training dataset, template for BanjoSet
- ../entities/basic-pitch — complementary: MIDI output but no string/fret
- ../concepts/musicxml-tab-notation — output format target