Algorithmic Music Generation With Conditioning

Date of Award

12-1-2022

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Computer Science

First Advisor

Proceso L. Fernandez, Jr., PhD

Abstract

This research seeks to create a basic implementation of an algorithmic music generator that uses the previous bar as a conditioning agent. Our proposed model, the Next Bar Predictor (NBP) is generally constructed using the three modules -- Data preparation, Operating-Training, and Generation. The output underwent initial listening evaluation by the developers. Models that successfully passed the initial evaluation proceeded to the second listening stage. Objective and subjective assessments were employed to compare the output of the different versions of NBP and the Baseline model, i.e., MidiNet. Objective assessments include the comparisons of the training times, dissimilarity score, and weighted averages of the out-of-scale notes. Subjective assessment on the other hand is a human evaluation test (19 novice and 11 professional listeners) using the three criteria: “how pleasing”, “how interesting” and “how realistic”.

Comparing the different versions of NBP and Midinet, NBP 2.1 recorded 9% of its notes to be out of scale.This is 39.2% better than the baseline model. The dissimilarity score, while lower than the baseline model, is nevertheless regarded as high with a score of 0.938. The simplicity of the NBP 2.1 implementation led to the architecture's efficiency, as evidenced by the overall training time, which is only 23% of MidiNet's total training time. Finally, based on human evaluations, the melodies created by NBP 2.1 are generally better than those generated from the baseline model. The paired t-test supports this conclusion, indicating that NBP 2.1 has a considerable advantage over MidiNet except for the interesting criteria where there is still an advantage for NBP 2.1, but such advantage is not as considerable as in the other criteria.

Link to Full Text

Share

COinS