eduzhai > Applied Sciences > Engineering >

Timbre latent space exploration and creative aspects

  • king
  • (0) Download
  • 20210507
  • Save

... pages left unread,continue reading

Document pages: 4 pages

Abstract: Recent studies show the ability of unsupervised models to learn invertibleaudio representations using Auto-Encoders. They enable high-quality soundsynthesis but a limited control since the latent spaces do not disentangletimbre properties. The emergence of disentangled representations was studied inVariational Auto-Encoders (VAEs), and has been applied to audio. Using anadditional perceptual regularization can align such latent representation withthe previously established multi-dimensional timbre spaces, while allowingcontinuous inference and synthesis. Alternatively, some specific soundattributes can be learned as control variables while unsupervised dimensionsaccount for the remaining features. New possibilities for timbre manipulationsare enabled with generative neural networks, although the exploration and thecreative use of their representations remain little. The following experimentsare led in cooperation with two composers and propose new creative directionsto explore latent sound synthesis of musical timbres, using specificallydesigned interfaces (Max MSP, Pure Data) or mappings for descriptor-basedsynthesis.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×