This page is about sound synthesis based on nonlinear dimensionality reduction, utilizing a method called UKR (Unsupervised Kernel Regression).

The basic idea behind our implementation is to treat a given recording as the basic material which is chopped to pieces, analysed and recombined by a certain model generated by UKR. Therefore the snippets extracted from the original recording are treated as a sample from a multivariate data distribution. Then, a principal manifold (a low dimensional representation and a continuous mapping between both spaces) is fitted to the data. New sounds are then created by moving along this manifold and playing back a mixture of all audio snippets lying in the neighbourhood of the selected position.

We implemented 3 different algorithms for mixing the snippets : One method based on granular synthesis, another utilizing the phase vocoder and the third method, that creates harmonic sounds from any kind of basic material by additive synthesis.

More detailed information will soon be posted here. At the moment, you have to make do with the demonstration software, which is available as a VST plugin (currently only for the Windows™ platform). To make use of this software, you need a VST compatible host, like Steinberg Cubase (VST/SX) or Emagic Logic Audio.

Download demonstration software

Sound examples

For a quick impression what UKR synthesis sounds like, you can also listen to the following examples available as mp3-files. The examples were created by fitting a 2 or 3-dimensional manifold to different sound material, using the above mentioned software. During synthesis, the position on the manifold was modulated by MIDI controller messages.