Participants 2023
muracle
AI Song Contest 2023 / participantsTEAM / muracle
SONG / Apocalypse
TEAM MEMBERS / jonghyun Kim
About the TEAM
As a person who likes to listen to various genres of music, I was thinking to make music through AI someday. I studied artificial intelligence in graduate school and now I’m building natural language processing (NLP) AI models in company. I personally have developed an AI piano composer model similar to language model and named the model as muracle (music + oracle).
I listen to the songs created by the AI model and select good songs and upload them to YouTube. In the future, I’m planning to develop an AI model that creates ensemble pieces beyond piano songs. I believe that someday AI will be able to create new music that goes beyond the level of human composition.
About the SONG
Apocalypse is a piano piece with fast tempo. The composition of this song (location / pitch / duration / velocity of each note) was all created by AI. AI has been learned through many piano pieces, and based on the previous notes predicted, it makes piano piece by statistically predicting the next note sequentially.
After that, the final song was made through post-processing, which added piano sustain pedal and lowered the piano's velocities only for parts that were too loud. It’s named Apocalypse because it is reminiscent of right before the end of the world and right after the apocalypse.
About the HUMAN-AI PROCESS
This song was mostly made by AI. The AI model is learned to create piano songs since piano music is easier to represent with discrete numbers than other genres and can express various emotions. The structure of the AI model is the same as the one used in natural language AI. It sequentially and statistically predicts the next note through the information of the previous notes.
AI models can produce many songs quickly, but all songs are not good to listen to. Therefore, I developed a score that can evaluate music numerically, and selected the top 1% of the songs out of thousands. The criteria for evaluating a song numerically are based on how similar patterns are to the training data and how little repetition is made in the song. Therefore, the evaluation prefers music that has a variety of melodies and has a smooth transition between them.
Automatically selected songs were exported to MIDI files containing predicted notes and listened through the MIDI player (Apple Garageband) to evaluate the songs by myself. I chose my favorite song among them and proceeded with the post-processing. The post-processing contains reducing the velocities of the parts where the sound is too loud and adding piano sustain pedal. After that I made the final song by selecting the sound font that fits the mood of the song.