Reproduce “MIDI-to-Tab: Guitar Tablature Inference via Masked Language Modeling” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Drew Edwards, Xavier Riley, Pedro Sarmento, Simon Dixon
A data-driven method for converting MIDI to guitar tablature employs a BART (Bidirectional and Auto-Regressive Transformers) model, framing the task as masked string assignment prediction. The system achieved 73.18% agreement with ground truth tablatures and was rated significantly more playable than commercial alternatives by experienced guitarists.
Key insights
- Deep learning models, specifically the BART architecture, can effectively learn implicit rules of guitar playability from large datasets, outperforming traditional rule-based methods.
- Framing guitar tablature generation as a masked language modeling task, akin to natural language processing, allows the application of Transformer models to musical sequence prediction.
- A two-phase training approach combining large-scale pre-training on crowd-sourced data with fine-tuning on smaller, high-quality professional data significantly improves performance.
- The novel quintile inference mechanism successfully addresses the training-inference asymmetry inherent in applying masked language models to sequential prediction tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.