Rhythm Guitar Tablature Continuation Through Picking Pattern Generation

Rhythm guitar tablatures in Western popular music often rely on repetitive patterns that follow a chord progression and usually preserve textural and rhythmic properties. This paper introduces a novel approach for suggesting rhythm guitar tablature continuations, conditioned on chord sequences and textural metrics. To benefit from a modular approach, we propose to split rhythm guitar tablature continuation into two steps: choosing chord positions and generating picking patterns. This paper focuses on the latter step, for which we propose two models: 1) a rule‑based model mimicking copy‑and‑paste practices in tablature notation software and 2) a Transformer‑based model that can generate more varied continuations using conditioning signals like textural variations and structural information. We train and test our models on the rhythm guitar tracks from the DadaGP tablatures dataset. The performance of the models is evaluated through several quantitative metrics as well as an online user study. While the Transformer model demonstrates better quantitative results, the rule‑based model is preferred in subjective ratings for its consistency and usability. Consequently, both models show promise: the rule‑based model for assisting tablature notation, and the Transformer model for providing creative suggestions during composition.