Postprocessing Technique for Algorithmically Composed Melodies

In a previous paper, we proposed the next bar predictor (NBP) model for algorithmically composing melodies. The NBP is much simpler than existing algorithms, yet it fared competitively with the best ones in all criteria (how realistic, how interesting, and how pleasing). In this paper, we propose a further improvement to the NBP by incorporating a melody post-processing technique that would generate a melody that better satisfies music theory and practice. The proposed technique ensures that all sufficiently long notes (based on a predefined threshold) align with a current chord while the shorter notes align with the given scale. As the proposed adjustment technique depends on the note duration threshold, we experimented with various thresholds and eventually decided to use quarter notes. We then compared the outputs of the standard NBP with the modified NBP (MNBP), and the results showed a significant improvement based on actual human evaluations involving music experts and enthusiasts alike. The result in MNBP improved by 9.33, 11.88, and 7.37% using the evaluation criteria of how realistic, how interesting, and how pleasing, respectively. A paired t-test using the 0.05 threshold shows that the improvement in the scores is statistically significant. Finally, since the post-processing technique can, in theory, be applied to the melody generated by any music composition algorithm, the algorithm was applied to Midinet, and surprisingly, even if the output of the Midinet is conditioned using the chord progression, results show that human evaluators prefer the adjusted melody of the Midinet than the original one, from the three criteria – realistic, interesting, and pleasing – the results improved by 20.82, 24.48, and 29.39% respectively. This is further validated to be statistically significant using the 0.05 threshold across all evaluators. Overall, the proposed technique does exhibit good potential for further improving algorithm- composed melodies.