Exploration and Mapping of Sound Synthesis Feature Vector Space

Exploration and Mapping of Sound Synthesis Feature Vector Space

Abstract

A report on recent work concerning the exploration and mapping of sound synthesizer feature space is presented. The feature vector space for different sound synthesis algorithms was investigated in various ways. Large scale maps of this space were generated, to provide an insight into the space of all possible sounds for different sound synthesis algorithms. First, a data set is generated by rendering many sounds from a sound synthesis algorithm, each with randomised settings. These sounds are then transformed into feature vector space using feature vectors such as MFCCs and perceptually weighted power spectra. This high-dimensional data set is translated to a low-dimensional space for human exploration using Multi-Dimensional Scaling (MDS). MDS aims to make a lower-dimensional map of a larger space whilst maintaining relative distances between points in both spaces. An interactive tool was developed which allows the rapid exploration of 2D MDS maps of feature vector space by a user.Different sound synthesis algorithms and audio feature vectors are compared using this tool. Finally, a fine grained analysis of the inter-point feature vector space is carried out in order to compare the smoothness of this space for different feature vectors.