Parametric optimisation techniques are compared in their abilities to elicit parameter settings for sound synthesis algorithms which cause them to emit sounds as similar as possible to target sounds. A hill climber, a genetic algorithm, a neural net and a data driven approach are compared. The error metric used is the Euclidean distance in MFCC feature space. This metric is justified on the basis of its success in previous work. The genetic algorithm offers the best results with the FM and subtractive test synthesizers but the hill climber and data driven approach also offer strong performance. The concept of sound synthesis error surfaces, allowing the detailed description of sound synthesis space, is introduced. The error surface for an FM synthesizer is described and suggestions are made as to the resolution required to effectively represent these surfaces. This information is used to inform future plans for algorithm improvements.
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member and would like to subscribe to the E-Library then Join the AES!
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.