Preliminary Investigations of Transformer Models for Bit-Domain Audio Synthesis and Timbre

I will be presenting on preliminary research findings of new work centered on the associations between Bit-Domain Audio Synthesis and timbre generation. Bit-Domain Audio Synthesis describes a class of techniques in which PCM audio samples are produced by closed-form integer expressions. These techniques include arithmetic and bitwise operations evaluated against an incrementing sample index. For example, the expression “t * (t >> 5 | t >> 8)”, in which the index t is combined with itself via right-shift, bitwise OR, and multiplication, produces sustained tonal material with evolving timbral character. This presentation reports on early experiments using compact autoregressive transformer models as discovery instruments for the relationship between expression structure and timbral outcome, trained on bit expression/audio pairs.

Speaker Biography

Daniel McKemie is a composer, technologist, researcher, and percussionist based in Oneonta, New York. His work focuses on audio feature extraction and low-level signal processing to explore timbre and novel synthesis techniques. Additionally, he designs innovative methods for interfacing handmade circuitry, controllers, and embedded systems with custom software to achieve his musical results. This approach enables the creation of complex, responsive performance environments and compositional processes controllable from multiple angles.

Daniel currently serves as a Lecturer at SUNY Oneonta, where he teaches courses on electronic music, popular music, percussion, and music psychology. He earned his MS in Computer Science at Brooklyn College, where he worked as a graduate research assistant in Professor Johanna Devaney’s Laboratory for Understanding Music and Audio (LUMaA), and served as an adjunct lecturer teaching courses in computer music. Daniel also has his MA in Music Composition from Mills College and BA in Music Performance from University of Nevada, Las Vegas.