Document Type


Publication Date



Deep neural network advancements have enabled machines to produce melodies emulating human-composed music. However, the implementation of such machines is costly in terms of resources. In this paper, we present NBP 2.0, a refinement of the previous model next bar predictor (NBP) with two notable improvements: first, transforming each training instance to anchor all the notes to its musical scale, and second, changing the model architecture itself. NBP 2.0 maintained its straightforward and lightweight implementation, which is an advantage over the baseline models. Improvements were assessed using quantitative and qualitative metrics and, based on the results, the improvements from these changes made are notable.