Teaching an Agent to Replicate Melodies by Listening : A Reinforcement Learning Approach to Generating Piano Rolls and Parameters of Physically Modeled Instruments from Target Audios

Detta är en Master-uppsats från KTH/Skolan för elektroteknik och datavetenskap (EECS)

Sammanfattning: Reinforcement learning has seen great improvements in recent years, with new frameworks and algorithms continually being developed. Some efforts have also been made to incorporate this method into music in various ways. In this project, the prospect of using reinforcement learning to make an agent learn to replicate a piece of music using a model of an instrument is explored. Both synthesizers and physically modeled instruments, in particular the Karplus-Strong algorithm, are considered. Two reward functions are introduced to measure the similarity between two audios: one based on frequency content and another based on waveform envelope. The results suggest that audio can be successfully replicated, both using a synthesizer and the Karplus-Strong algorithm. Further research can be conducted on replicating more complex melodies and creatively composing using physical models of instruments. https://github.com/wille-eriksson/RL-instruments

  HÄR KAN DU HÄMTA UPPSATSEN I FULLTEXT. (följ länken till nästa sida)