Research Aptitude Defence

My Research Aptitude Defence, the first major milestone of the PhD program at Dalhousie’s Faculty of Computer Science, took place on Friday February the 16th, 2024.

After a presentation to my committee and a larger audience, two rounds of questions, and deliberation, I successfully passed my RAD! I’ll now look onwards to the next deadline, the thesis proposal defence.

Title

Towards a Generative Machine Learning System for Effective Synthesizer Patch Creation

Abstract

Synthesizers are a popular type of musical instrument capable of producing a wide range of different sounds thanks to their user-adjustable parameters. By tweaking these parameters, musicians can precisely shape the sound of their instrument and save the result as a patch. This research at the intersection of Human Computer Interaction (HCI) and Machine Learning (ML) aims to explore how generative ML architectures, such as Variational Autoencoders, can be used to support musicians in generating new synthesizer patches using approachable tools. Using the open-source synthesizer amSynth as a test bed, we show how users can generate many previously unseen patches via interpolation of a low-dimensional latent space. Our initial experiments show that despite a lossy reconstruction of synthesizer patches, there is a high level of quality and diversity found in the synthesizer patches generated by our system. Furthermore, we implement bindings between FAUST synthesizers and the libmapper project in parallel with our work on generative ML in an effort towards making our results generalizable to additional synthesis paradigms and synthesizer form-factors.

This Research Aptitude Defence provides an extended summary highlighting the research conducted to date and published at an international peer-reviewed conference. This summary includes a discussion around the development of a ML system for synthesizer patch generation using Variational Autoencoders as well as the introduction of latent representations for synthesizer patches, namely “Latent Coordinates” and “Timbral Representation”, both of which are derived through analysis of the learned latent space. Furthermore, a report on the continued development of mapping tools for digital musical instruments, specifically targeting instruments written in FAUST, is also presented. Finally, this report concludes with a discussion regarding the road-map for future work with regards to this research.