Pay (Cross) Attention to the Melody: Curriculum Masking for Single-Encoder Melodic Harmonization

Konstantinos Theodoros Tsamis, Dimos Makris, Konstantinos Soiledis, Vassilis Katsouros, Maximos Kaliakatsos-Papakostas

Keywords: Symbolic Music Processing, Music Harmonization, Automatic Chord Recognition, Transformer Models, Single Encoder Architectures, Masked Modeling, Curriculum Learning, Music Information Retrieval, Sequence Modeling

Domains: Other

TL;DR: A single encoder transformer framework for symbolic music harmonization and chord recognition using progressive masked curriculum learning.

Abstract:

Symbolic music understanding remains a challenging problem in computational music research, requiring models to capture complex temporal, melodic, and harmonic dependencies across musical sequences. Tasks such as music harmonization and automatic chord recognition are examples of these challenges, as they require both local musical consistency and long-range structural understanding.

To address these challenges, we explore the use of a Single Encoder (SE) architecture for symbolic music harmonization. Unlike conventional encoder–decoder approaches, the SE framework formulates harmonization as a masked sequence prediction problem using a unified transformer encoder, where melody and harmony representations are processed jointly within a shared sequence representation. This allows the model to learn contextual musical dependencies while generating harmonically coherent outputs within a single latent space.

To further strengthen the dependency between melody and harmony, we employ a Full-to-Full masking curriculum during training. During the early stages, the harmony sequence is fully masked, forcing the model to rely entirely on the melodic input. As training progresses, increasing numbers of harmony tokens are gradually revealed until the sequence becomes fully unmasked, encouraging stronger melody-conditioned harmonic learning and improved structural consistency.

Published in the proceedings of the IEEE BigData 2025 3rd Workshop on AI Music Generation (AIMG 2025).

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 147