There is something artistic in the introspection required for understanding mathematical abstraction. As there is something analytical when you listen to a piece of music. This is what makes the research field of Sound and Music Computing (SMC) so interesting to me!
Thesis in completion to the title of Master in Music Production Technology and Innovation at the Berklee College of Music.
Abstract
In recent years, the development of generative AI models that are able to generate high quality audio has opened a new world of opportunities in the field of sound synthesis for musical applications. The possibilities inaugurated by this emerging technology, however, are only beginning to be investigated. In this work, Gen-Synth, a synthesizer combining generative timbre transfer and granular synthesis is proposed in order to further explore the artistic implications of AI in sound design. The work was developed by training a popular neural synthesis model - RAVE (Realtime Audio Variational autoEncoder) - with a library composed of Brazilian percussion recordings. Later, the trained model was used to develop a Max for Live device that combined a timber transfer module with a granular synthesis module. Finally, some user testing was performed, and feedback was gathered for future improvement.
Thesis in completion to the title of Master in Electrical Engineering at the State University of Campinas (UNICAMP).
Abstract
Multitrack mixing is the process in which the various tracks of a musical piece are processed and combined to achieve certain technical and artistic goals. As this is a task of high complexity and subjectivity, in recent years, artificial intelligence applications that seek to fully automate mixing have emerged. This study aimed to review two of the most recent artificial intelligence models proposed for this task and subsequently measure their results through a subjective preference test, where the mixes made by the models were compared with each other and also with mixes made by humans. Our results indicated that human mixes were preferred most of the time by the participants, regardless of each subject's level of mixing knowledge. When the mixes of the two models were compared with each other, the results diverged, sometimes indicating a preference for one model and sometimes for the other. Some possible areas of improvement and future directions were identified, as well as ethical issues regarding the development of this class of models
Thesis in completion to the title of Bachelor in Electrical Engineering at the State University of Campinas (UNICAMP).