Multi-task learning-based temporal pattern matching network for guitar tablature transcription

Abstract

Guitar tablature transcription poses unique challenges in automatic music transcription, as it requires capturing both pitch and string usage on a multi-string instrument with various expressive techniques. While guitar tablature is widely used by guitarists in the music field, neural architecture modeling for this task remains underexplored, particularly in accurately mapping pitches to their respective strings. In this work, we propose a multi-task learning-based temporal pattern-matching network (TPMNet) that effectively captures temporal information from guitar recordings, improving the alignment of predicted results. The key contribution of this work is the advancement of neural network architecture, leading to notable improvements in prediction performance for guitar tablature transcription. Additionally, we explored the optimal pooling layer selection method tailored to different tasks, addressing a long-confusing problem in the field. TPMNet’s efficacy was validated through experiments on the GuitarSet dataset, and its generalizability was confirmed via cross-evaluation with the EGDB dataset.