Reproduce “GAPS: A Large and Diverse Classical Guitar Dataset and Benchmark Transcription Model” on OpenResearch

About this paper

Xavier Riley, Zixun Guo, Drew Edwards, Simon Dixon

GAPS introduces a 14-hour, publicly assembled corpus of real classical-guitar performances with aligned audio, MIDI, MusicXML scores, tablature, and video metadata spanning 300 pieces, 205 performers, and 93 composers. A high-resolution transcription model trained on the dataset reaches 88.1% onset F-measure on GuitarSet without GuitarSet training and up to 94.3% on GAPS, indicating that performer and recording diversity can improve cross-dataset generalization.

Key insights

• Variation across performers, guitars, microphones, and recording environments can improve transfer to an unseen guitar dataset, even when the target uses different guitar and pickup types.
• Piano-transcription pretraining, especially from an augmented MAESTRO checkpoint, consistently outperforms training the guitar model from scratch.
• The dataset’s aligned scores and tablature extend its use beyond pitch detection to score following, performance analysis, fingering studies, expressive timing, and multimodal guitar research.

Summary by alphaXiv.