GAPS: A Large and Diverse Classical Guitar Dataset and Benchmark Transcription Model
Xavier Riley, Zixun Guo, Drew Edwards and Simon Dixon
Centre for Digital Music
Queen Mary University of London
ISMIR 2024
This work was supported by the UKRI and EPSRC under grant EP/S022694/1
Motivation
Piano transcription works really well (98% on MAESTRO, 90% on MAPS)
Why? Because there are good datasets for piano.
What happens if we build an equivalent dataset for guitar?
ISMIR 2024
Credit: GiantMIDI-Piano Kong et al. 2020
GAPS
300 Pieces (14 hours) of solo classical guitar with:
All manually checked and verified
Dataset
(Guitar-Aligned Performance Scores)
ISMIR 2024
GAPS
Training Data Pipeline
ISMIR 2024
GAPS
Results
ISMIR 2024
Method | F-measure (no offset) |
MT3 (2022) | 90.0% |
Lu et al. (2023) | 91.1.% |
Ours | 91.2% |
GuitarSet – test split
Method | F-measure (no offset) |
MT3 (2022) | 32.0% |
Zang et al. (2023) | 70.2% |
Maman and Bermano (2022) | 82.9% |
Ours | 88.1% |
GuitarSet – entire dataset (zero shot)
GAPS
Thank you!
ISMIR 2024