Multi-Band Audio Mosaicing

April 2020

project description:

The aim here is to reconstruct a given sound (referred to as "target sound") using time-sequence frames from other various sounds (or "source sounds"), all downloaded from the collaborative repository Freesound. The synthesis technique behind this process is referred to as "audio moscaicing" and consists in reconstructing a given audio time- sequence out of fragments of other sounds. Although the main intuition behind this would be to reconstruct the original signal with similar fragments, we can opt to pick a more creative path by defining other reconstruction criteria. In this work, we opt to choose a rather experimental path by reconstructing our signal on a per-band basis. We call this multi-band audio moscaicing. The idea behind this experiment is to split our target sound (that is, the sound we are trying to reconstruct), into three different spectral bands namely "low", "mid", and "high". Once this process is done, each of the spectral band is treated as its own target file. The figure below depicts the overall workflow behind this process.


Below are some example of such process applied to a full-band jazz recording:


Target Audio



00:00 / 00:11
00:00 / 00:11
00:00 / 00:11
Reconstructed Audio

© 2020 Darius Petermann