Statistical concatenative synthesis

Statconcat was completed as my undergraduate disertation. It looked at the use of a markov chain as part of a unit selection algorithm for concatenative synthesis. The project is made up of a Pure Data patch with some Java externals.

Usage

  1. Under "Search" click "Folder". Select a folder containing the sound files you would like to use to populate the "Database". Then click "Start".
  2. The selected files will then be sliced, analyzed and stored.
  3. Under "Input" "Target" click "L". Select a synthesis target that will drive the unit selection.
  4. If you check the pure data console you should see a starting message. After a few seconds the output buffer should be populated with samples. Click "P" to play the re-synthesized sound.

Issues

This is a very old project and was one of the first programs I had written in Java that actually did something useful. As such, there are many problems and the code is absolutely awful.

The patch does not handle large samples very well, so it is best to use short samples (under 30 seconds or so) and avoid loading hundreds of files. There is also a problem where the same target sound cannot be loaded consecutively, otherwise the patch will crash.

Apart from that it may still be useful, and it is certainly possible to get some weird sounds out of it.

Example

Here is a quick audio example. The sounds that were used for resynthesis come from the usual drum breaks (amen, funky drummer etc.) and some short reggae vocal samples.

Target sound:

Resynthesised sound using the markov chain:

Code can be found on github