A Comprehensive Study On Music Generation Using A Generative Adversarial Network Conditioned On Mood

The omnipresence of music in nowadays' everyday life renders the creation of music a process of high demand.The entertainment industry requires the constant production of novel songs and melodies not only to publish music per se,but also as background track for movies,television,theater and countless more.In order for musicians to provide this immense amount of novel musical pieces,their process needs to be efficient in both time and expenses.Furthermore is the practice of making music not only a profession,but one of the most popular hobbies people pursue in modern society.While music can be a solo pastime,with the rise of the internet elements like play alongs for improvisation or automatic background generation to replace a band have taken root in many hobby musicians' routines.Like in many areas,a bottleneck of both efficiency and cost is the limitation in human resources.Modern computer technology enables society to automate an ever growing variety of tasks with the same previously known limitations.Music generation is a highly complex process compared to the current state of research,and therefore cannot be achieved by basic automation algorithms.As shown in related work,a much more appropriate approach is the application of artificial intelligence,more precisely deep learning.In this work we incorporate the concept of progressively growing generative adversarial networks to generate music.We furthermore enhance the mere generation by making the generative process conditional,enabling the input of musical moods as a parameter to consider in the generation process.This way,our resulting model achieves novel functionality in this area.Concludingly,in this work we achieve and present the following contributions.1.Data In order to train our C-Midi PGAN,we create a new dataset containing roughly 50000 musical samples of 10 seconds each.These samples are gathered by splitting 1000 classical songs in MIDI format written for piano into the 10 second sub-samples.The MIDI data is web-scraped.2.Method We successfully create a progressively growing generative adversarial network by state of the art design.This model is focused on generating MIDI data and accordingly enhances comparable state of the art.3.Advanced method We enhance the baseline model to be conditional and therefore create a novel model.To sensibly provide musical mood as a condition,we introduce a way to label our dataset in a scalable and arithmetic manner.This enables highly flexible fine tuning of the generated samples and serves the very specific and detailed needs of musicians.4.Application system Our full model is made available in a user friendly web environment in the form of an SPA.This application offers basic features to work with our method,as well as the potential to gather useful data from users in order to be evaluated in the future.5.Verification We finally evaluate our results by means of objective and subjective metrics,scoring an FID score of 12.30,rivaling the state of the art,as well as conducting an extensive user study which renders our model comparable to related work in some aspects,and better than related work in other aspects.The contributions presented in this work further the abilities of state of the art generative neural networks by improving the task of music generation.Our C-MidiPGAN introduces the novel feature of conditional music generation by mood.The dataset created for this project bears usefulness for future work in the area of music generation.Our web application offers an efficient way of not only gathering data,but also improving the presented method constantly.