TabCNN

About

CNN-based guitar tablature transcription from audio. Takes CQT spectrogram input and outputs a 6x21 (strings x frets) representation — directly predicting which string and fret produces each note. Trained on GuitarSet. This is one of very few models that outputs tablature rather than MIDI pitches. Code is minimal (no ready-to-run package).

Relevance

Most directly relevant architecture for banjo tab transcription. Adapting to banjo requires changing the output from 6x21 to 5x22 (5 strings, ~22 frets) and training on banjo data. The approach — CQT spectrogram to stringxfret CNN — is directly transferable. The main blocker is the lack of banjo training data (GuitarSet equivalent).

Mentions

• ../sources/2019-11-04-tabcnn — paper
• ../entities/guitarset — training dataset
• ../sources/2024-04-19-synthtab — SynthTab, successor using synthetic data

Links

• GitHub: https://github.com/simon-minami/tabcnn (minimal)
• Related: https://github.com/cwitkowitz/guitar-transcription-with-inhibition (improved version)