ABOUT THE TEAM

We are a committed team who lives in La Paz and studies at Universidad Privada Boliviana. Most of us met in the first year in college studying the same career: Computer System Engineering. Since then, we have worked together many times consequently and participated in contests and projects such as Cybersecurity Competition, ICPC or being part of IEEE Student Branch. At this time, in our fourth year in college, we matched in the same class of Artificial Intelligence (AI), so we made this team to make one of the projects related to this subject. As we chose the music generator, our teacher motivated us to participate in this AI Song Contest.

ABOUT THE SONG

In the beginning of a common day we afront different challenges to achieve our goals and personal victories. The same thing applies when a character of a videogame starts its journey, and we believe this song makes us feel like this is the beginning of this particular character, which is an algorithm who was recently born in this human world. Why is it here? What is it meant to do? So it begins its voyage through our society in order to learn who it is. This song was created with a combination between us and Artificial Intelligence. First, the AI composed different melodies and rhythms, then we selected the best ones to mix them and create an interesting song.

ABOUT THE HUMAN-AI CO-CREATION PROCESS

This song was composed using melodies from MusicVAE, a machine learning model which let us create scores and melodies that later can be mixed in a process called “Interpolation”. This interpolation process receives 2 melodies, the idea is to generate a gradual “road” from the first one to the second one.

The creation using AI started selecting different Datasets such as Maestro, VGMIDI or, NESMIDI, most of them related to videogames. Then, the model was trained using a predetermined configuration specialized in piano songs, called “cat-mel_2bar_big”. This part was the longest one, not just because the model had to be trained enough, but also because it should not overfit. Once the model was ready, the next step was to generate different kinds of examples to evaluate and select which ones were the best in order to interpolate and mix them in a single composition.

Finally, with the best examples selected and the interpolation process we composed the resulting song.