Muusika saate genereerimine tingimusliku vastandgeneratiivse närvivõrgu abil

Abstract

Generation of good quality music accompaniment is very useful for composers and music producers. Generative artificial neural networks are booming and there is recently an increasing amount of music generation models published. The aim of this Bachelor’s thesis is to generate music accompaniment using spectrograms and an image translation model Pix2Pix. Experiments are con-ducted to generate different types of accompaniments. The best results are achieved when gener-ating the drum stem. It can be seen from the results that generative adversarial networks’ outputs contain unnatural artifacts that affect the results badly. Preventing this requires lots of finetuning.