Abstract
I will be presenting an ongoing research collaboration (with Dr. B. Sturm, KTH) using deep learning on melodic transcriptions. Our initial data set is a large (>20,000) collection of folk tunes crowd sources online. We used those to create a model that produces plausible outputs that share many of the characteristics of the original style ( folkrnn.org) . But as a composer, I wanted to prod this model in less conventional directions. Doing this, we discovered that the system did not in fact learn what we initially thought it did. My talk will explain our approach to machine learning with focus on evaluation of the model and the link between computational and creative research.
Bio
Oded Ben-Tal is a composer and researcher working at the intersection of music, computing, and cognition. His composition include both acoustic pieces, works combining instruments with electronics and multimedia work. Much of his recent composition work focuses on techniques borrowed from machine listening research for interaction between performers and computers. His compositions have been performed around the world including Italy, The US, Korea, Denmark, Isreal and the UK. He is a senior lecturer at the Music Department, Kingston University.