UC San Diego

Cross-synthesis is the process of blending two or more audio signals to produce a hybrid signal with timbral characteristics of the originals. There are a number of methods to perform cross-synthesis of digital audio including but not limited to vocoding, phase vocoding and convolution. Convolution is a common approach to perform cross-synthesis on arbitrary sound files but the process suffers from inflexibility. Having no control over the process itself, a musician interested in convolutional cross-synthesis is left to modify the original sounds to influence the outcome.

There are a number of drawbacks with the process of convolutional cross-synthesis that impede its musical usefulness. One such drawback is the lack of control over which sound the outcome more closely resembles. Another is a perceived absence of high frequency energy in the outcome. This thesis will introduce novel extensions to the process of discrete convolution for the purposes of offline cross-synthesis that attempt to remedy these concerns.

Discrete convolution and the proposed extensions will be analyzed for their effect on acoustic features. This black-box analysis consists of subjecting a novel, heterogeneous dataset of sounds dubbed RFS1k to convolutional cross-synthesis and examining the average effects on a set of features.