AbstractA guitarist's search for `tone' -- their ideal timbre -- can lead them to exploring countless guitar effects. Many players yearn for the tone of equipment from the early days of electric guitars. Now quickly becoming a mature field, Virtual Analogue modelling aims to digitally emulate analogue audio equipment in real-time, making potentially rare devices more accessible. A subset of Virtual Analogue of particular interest is found in the physical modelling of audio circuits, where circuit-level models are built from models of electronic components, drawing upon both widely transferable physical concepts and engineering methods.
Despite the increasing ubiquity of physical models, the simulated input/output behaviour is rarely compared to that of real circuits. The main contribution of this work is to reconcile this disparity with the presentation of two complementary identification procedures that aim to find a model with minimal difference to a reference circuit. Focusing on guitar pedals, measurements of the circuit are taken solely from existing input and output connections to reduce the required measurement time in comparison to measuring each component individually, which also prevents any damage being caused from the deconstruction of the device.
The identification approaches proposed in this study utilise an optimisation algorithm that minimises the difference between the output of candidate models to that of a reference circuit by modifying the values of the physical component parameters. Within the required simulation, the solving of nonlinear equations is a likely source of inefficiency and even failure, prompting the search for an algorithm that avoids these issues. Uncertainty about the accuracy of less well understood components can also lead to difficulties in the circuit identification. A component that is found to be markedly different is the germanium BJT -- a core component present in a vintage case study -- and is thus the focus of a component-level identification.
Of the two proposed identification procedures, the first aims only to minimise the output error, discarding accuracy at a component level, and placing a focus on minimising the computational expense of the identification. In addition to high fidelity models, results point towards a strategy to overcome the curse of dimensionality when addressing circuits with a large number of components. The second, more physically valid procedure aims to retrieve accurate parameter values of each of the circuit's components such that the estimated component values remain valid under modifications to the circuit. To address possible non-convergence problems, an approach is developed that makes use of multiple measurement sets involving additional components of known value, thus introducing further constraints on the search space. The performance of both procedures is exemplified and evaluated by means of case studies.
|Date of Award||May 2019|
|Supervisor||Maarten Van Walstijn (Supervisor) & Robin Ferguson (Supervisor)|