The ACIDS research group is dedicated to crafting neural audio synthesis tools for musical production and performance. Over the past years, the group has developed a range of technologies for timbre transfer, controllable instrument generation, network compression, analysis, and model bending.
This presentation introduces the underlying principles and motivations behind these developments, bridging technical, musical, and scientific perspectives. It explores the new tools and creative possibilities opened up by neural audio synthesis, while presenting recent updates to existing systems and highlighting the latest research and artistic collaborations that have led to new approaches.

Latest collaborations : Canblaster & Neurotipyque @ Marathon festival, Pierce Warnecke @ Ircam Forum Session, Molécule @ Ircam.
Technologies
Ravetable ~ What if latent signals were treated as audio signals? Ravetable explores this idea to build a latent-based wavetable synthesizer built on top of the RAVE model, enabling time-synchronised, zero-latency neural audio synthesis.

AFTER ~ AFTER combines the versatility and expressivity of latent based neural synthesis with accurate control over known musical features such as melodic content. This presention will unfold the latest developement to the AFTER system and Max4Live device, with new synthesis paradigms such as descriptor based synthesis and bpm clock conditonned drum break generation.

TorchBend ~ Inspired by circuit bending, Torchbend allows to manipulate the internal representations learned by neural networks in order to deviate, push, or break the synthesis process and create entirely new soundscapes.
JUNK ~ Junk is real-time hardware instrument built around a Raspberry Pi with dedicated MIDI controls that can host nn~ models such as RAVE.