Michael S. Lewicki
Richard M. Stern
David Wessel (University of California, Berkeley)
Abstract:
I propose an approach to create high-quality music synthesis by automatically constructing an instrument model and a performance model; the latter module generates control signals from score input, and drives the former module to produce synthetic instrumental sounds. By designing and applying appropriate machine learning techniques, the instrument model and the performance model can be trained with acoustic examples and their corresponding scores for a musical instrument. The automated model should be superior in performance to one manually modeled and tuned by hand.
In this proposal, I describe the framework of the automated synthesis and modeling system, explain the reasons for employing specific techniques, such as the necessity of modeling amplitude envelopes and a machine learning approach, discuss specific characteristics of individual modules, and propose several possible solutions for some parts of the system.
I state the contributions of the thesis topic, define criteria for project success, explore possible future work, and present a timetable for the research. I am confident that the thesis topic will yield interesting results and believe I can finish the proposed tasks within the scheduled time frame.
