This thesis examines how assistive Artificial Intelligence (AI) can be integrated into sound design workflows, lowering the learning curve for novice sound designers while enhancing workflows for advanced sound designers. This is accomplished through the creation of Tensynth—an AI-assisted software synthesizer whose name reflects the use of “Tensors”, multi-dimensional arrays that form the backbone of modern AI systems. Tensynth generates fully editable synth patches from audio examples, as Audio-to-Parameters, automatically mapping audio references to a set of 69 parameters, or natural-language descriptions: Text-to-Parameters, automatically mapping semantic text to over 100 parameters.
Recent advances in neural audio synthesis have produced impressive generative models; however, they often lack usable user interfaces, practical workflows, speed, transparency, and necessary user control. To address these gaps, this thesis uses a mixed-methodology approach, combining practice-based research methods that focus on user-centred design frameworks to reveal and reduce inefficiencies in current sound design workflows.
Tensynth uniquely maps user input to synth parameters with real-time AI inference, visual feedback, and patch-aware MIDI auditioning: a feature that matches MIDI patterns to the specific characteristics of the generated sound patch, providing users with instant feedback and understanding how to use their results in context. Comparisons with existing sound design workflows and user testing demonstrate that Tensynth lowers the novice learning curve for recreating sounds and enhances genuine creativity and artistic intent. As a result, Tensynth serves as a model for the next generation of intelligent, accessible, efficient, and editable systems that support learnability, creativity, and expert craft, not just in music production but across all domains that involve complex AI parameter mapping.