Generative Music: Seq2Seq Models for polyphonic enrichment

Abstract

Music generation aspires to produce musical pieces that have a pleasant harmonic result along with a complex and stimulating context, engaging listeners in emotional, intellectual and aesthetic dimensions. Polyphonic Music Generation can be modeled as multi-dimensional language generation task, thus, the contribution of state-of-the-art models, techniques and approaches in Natural Language Processing can be more than beneficial when applied in musical context. In this thesis, we propose a Seq2Seq generation approach, designed to enhance compositions by integrating additional voices, incorporating either an upper or lower voice to the given melody. To transform symbolic music contained on a MIDI file into sequential format, suitable for transformer models, we employ modern encoding tools. This enables us to integrate advantages such as byte pair encoding, data augmentation, and other techniques commonly used in Natural Language Processing. Despite music evaluation being a subject of disagreement among members of the research community, the field remains active for academic exploration. In our evaluation process, we introduce a novel approach that integrates quantitative assessment with qualitative criteria, aiming to bridge the gap between objective metrics and human perception