phase 1 phase 2 - guitar.ucsd.edu
TRANSCRIPT
Satvik Nagpal, Conan Lu, Elena Atluri
ABSTRACTOur project aims to convert music into a given style using machine learning. We accomplished this using an algorithm known as a generative adversarial network (GAN). A GAN consists of a generator, which creates music based on an input, and a discriminator, which judges if the generated music is real or fake. We fed in pairs of pitch classes and piano rolls to train the algorithm in a given style. Then, the trained generator can be used to interpret other musical pieces into the style it was trained on.
Example Our Project
OBJECTIVES1. Generate pop music using
machine learning.2. Convert a piece in one style to
another style (in this case, pop).
ACKNOWLEDGEMENTSProfessor Shlomo Dubnov, Professor Mauricio de Oliveira, Jacob Sundstrom, Aren Akian, Gualter Moura
CONCLUSIONGiven the success of the test demo (Pachelbel's canon), our hypothesis that GANs can be utilized to convert music into different genres is somewhat supported. In order to truly test our hypothesis, we need to run our program with a larger dataset that includes different music styles. To develop our product in the future, we would have to train the computer extensively and manually adjust the GAN algorithm to match our needs.
HYPOTHESISWe hypothesize that we can utilize the GAN model to convert pieces from one style to another, by training the network to emulate the attributes of a specific style.
HOW GENERATIVE ADVERSARIAL NETWORKS WORK
PHASE 1 PHASE 2TRAINING THE GAN - Successfully generate music using chromas
USE THE GENERATOR - Convert music in other
genres to pop
WHAT IS A CHROMA?Chromas encapsulate the chords and general prevalence of notes in a song. Chroma puts every note on a 12 value spectrum to provide a visual representation of the notes in a musical piece.
TRAINING DETAILS122 midi files (piano scores) of pop and their chromas, as pairs, are used to train the model.
GENERATED PRODUCTAfter 357,000 epochs (training iterations), the computer outputted this.
OUR THEORYGANs are popular for image classification. In the same way, GANs can be used to classify and generate music.
GENERATED PRODUCTDEMO INPUT (Canon in D)https://soundcloud.com/conan-lu/original?in=conan-lu/sets/ganmidiOUTPUT https://soundcloud.com/conan-lu/out?in=conan-lu/sets/ganmidi