Generative Modeling of Bach-Style Symbolic Music: A Comparative Study of Autoregressive, Latent-Variable, and Adversarial Approaches

A comparative study of three generative-model families for Bach-style symbolic piano music finds that autoregressive LSTMs with attention produce the most coherent samples, while vector-quantized variational autoencoders mitigate a common training failure and adversarial networks struggle to generalize reliably to the composer’s style [1][2]. The work, posted on the arXiv preprint server, evaluates autoregressive, latent-variable, and adversarial approaches on a shared MIDI corpus of polyphonic note sequences [1][2]. The autoregressive LSTM with attention yielded the most musically coherent compositions, according to the authors’ experiments [2]. Vector quantization helped avoid posterior collapse — a condition where the latent representation is ignored — and generated more structured outputs than conventional recurrent VAEs [2]. The generative adversarial network captured local pitch patterns but remained difficult to train and generalized less reliably to Bach’s idiom [2]. The study was submitted by Dezhi Yu on 11 June 2026 and revised the following day [1]. It appears under the Sound category of arXiv, an open-access repository that hosts electronic preprints across disciplines including computer science, physics, and mathematics [6]. arXiv was launched in 1991 and by late 2021 had surpassed two million articles; as of November 2024 the submission rate stood at roughly 24,000 papers per month [6]. The repository is moderated but does not conduct peer review [6]. The paper’s landing page integrates several community-developed tools through arXivLabs, a framework that lets third parties build experimental features on top of the repository [4]. Available services include the Bibliographic Explorer for navigating citation trees, the CORE Recommender for surfacing related open-access papers, and Connected Papers for visual literature mapping [5]. arXivLabs collaborators must adhere to the repository’s values of openness, community, excellence, and user-data privacy, and are granted only minimal, anonymized data necessary for their features to function [4]. The comparative results highlight distinct failure modes across model families and add to a growing body of machine-learning research on symbolic music generation [2]. The authors frame the work as a systematic benchmark rather than a claim of a single best method, noting that each architecture exhibits relative strengths depending on the evaluation criterion [2].