Demixing and Remixing Music with Deep Learning

Abstract: In 2015 Alejandro Koretzky created tuneSplit with the goal of democratizing music creation and remixing while introducing the concept of “Semantic Equalization” in music. By implementing an end-to-end pipeline that performs audio source separation in real time, commercial stereo music can be deconstructed into different instruments and vocals, allowing users to personalize the listening experience and unlocking the possibilities for remixing using parts of existing stereo mixes. Initial versions of the underlying algorithms were based on a proprietary adaptive version of Non-negative Matrix Factorization. Koretzky's latest work based on Convolutional Neural Networks (CNN) achieves state of the art results in real time with significant quality improvements over the previous techniques.

Authors: Alejandro Koretzky, Karthiek Reddy Bokka, Naveen Sasalu-Rajashekharappa

Speaker's Bio:  Alejandro Koretzky was born in Córdoba, Argentina, where he received his Bsc. in Telecommunications Engineering. He held positions as a Software Engineer, Data Scientist and Product Manager at different companies such as Motorola Solutions and co-founded various ventures including tuquejasuma.com. In 2013 he came to the U.S. with a Fulbright Presidential Fellowship to pursue a Msc. in Electrical Engineering at USC, with a focus on Signal Processing and Machine Learning. In 2014 he started a research project under the supervision of Prof. C.C. Jay Kuo, which later evolved into the first prototype for tuneSplit. tuneSplit was incubated by the USC Viterbi Startup Garage program and selected for a grant from the National Science Foundation's I-CORPS program. In 2016 it was acquired by Red Pill VR, a virtual reality startup working on the future of social music experiences in VR. Alejandro is also a singer, songwriter and composer, having released two studio albums, with songs featured on MTV and VH1 Argentina.