Date Approved

9-4-2026

Graduate Degree Type

Thesis

Degree Name

Applied Computer Science (M.S.)

Degree Program

College of Computing

First Advisor

Haoyu Li

Second Advisor

Jiaxin Du

Third Advisor

Yong Zhoung

Academic Year

2025/2026

Abstract

Systems that accompany a live musician, such as ACCompanion, work by retrieval: they hold a written accompaniment and stretch its timing to follow the soloist, so they can only accompany music whose second part already exists as a score. This thesis asks whether that second part can instead be generated, and specifically whether a machine learning model can write an accompaniment for a piece whose score it has never seen. A player performs the primo, the melody; a model produces the secondo, the accompaniment; and a score follower places the generated notes against the live performance. If this works, a written second part is no longer required in order to play with an accompanist.

The question is answered in stages, each deciding whether the next is worth attempting. A pretrained music transformer is first run untouched, to establish what a general model already does: it produces a coherent accompaniment in the right register, but it writes for an ensemble rather than for a single keyboard, and is not idiomatic. Fine tuning is then applied to a single piece with a low rank adapter, a small set of extra weights added to a frozen model, deliberately overfit. The generated secondo comes out essentially identical to the written score, note for note, and played live against recordings of an expert human duo it matches that written score under the same follower. This establishes the central possibility, that an accompaniment of score quality can be generated rather than retrieved, and it shows that fine tuning is what converts the base model's raw ability into that quality.

Fine tuning on a single piece memorizes it, so the final experiment puts the real question to a wide corpus. An adapter trained on the Bach four part chorales is asked to accompany chorales it never saw. Because an unseen melody has no single correct accompaniment, closeness to one reference cannot measure quality; the generated accompaniments are judged instead by Frechet Music Distance, which compares the generated set and a set of real Bach as distributions and therefore measures authenticity rather than agreement. The model trained on many chorales places its output measurably closer to real Bach than one trained on a single chorale, on pieces neither had seen. The result is preliminary, resting on one corpus and a held out validation set, but it is direct evidence that an accompaniment for an unseen score can be generated by machine learning, and that breadth of training is what makes it possible.

Download

Share

COinS