9:40 - 10:00
Yifei Teng, An Zhao
University of Illinois at Urbana-Champaign
Machine Composition of Pop and Rock Music
We present a hybrid neural network and rule-based system for composition of pop and rock music. A temporal production grammar was used to generate chord progressions and musical structure, while a conditional variational autoencoder was trained to fill in the melody. Techniques are proposed to identify the melodic track from the many tracks in a MIDI file and to determine the chord played at each time step. Over ten thousand MIDI files have been analyzed and segmented into training samples lasting around four to eight measures each. The melody was used to train the autoencoder, which maps them into a 64-channel feature space, conditioned by the underlying chord progression. The system generates melodies by feeding a random sample taken from that space into the decoder, along with a chord progression produced by the grammar. Outputs compare favorably against other commercial and academic solutions.
Design downloaded from free website templates.