Latent Sampling: a PLAUD Performance

This performance presents a live set built around PLAUD, a DDSP architecture which encodes corpus timbres into a variational latent space and learns its temporal character via transformer-based autoregressive priors. The performance is built around the use of multiple specialized models trained on small, curated corpora. Trajectory sampling is a central interaction mode that translates familiar audio sampling operations into the latent domain: control paths from manual modulation or the prior can be looped, sliced, reversed, and replayed at variable speeds. Differently from audio sampling, due to the temporal context, the synthesised sound depends not only on the punctual playback position, but also on the broader playback conditions, including the preceding control trajectory and playback speed. Bending operations further intervene via component limiting and waveshaping, producing out-of-distribution behaviour. The performance attempts to expand the gestural repertoire of audio sampling by leveraging the material affordances of the neural synthesis.