Neural Resonator VST

Generate and use filters based on arbitrary 2D shapes and materials

Demo video 🎥

About

This is a VST plugin that uses a neural network to generate filters based on arbitrary 2D shapes and materials. It is possible to use midi to trigger simple impulses to excite these filters. Additionally any audio signal can be used as input to the filters.

This plugin uses LibTorch for the neural network and JUCE for the VST plugin. It also features an easily extensible web-based user interface built with React and Typescript. However, it is possible to build the plugin with a traditional C++ GUI (Linux will use this as default).

See also the implementation of this project on the Bela platform:

https://github.com/rodrigodzf/NeuralResonatorBela

Installation

The plugin is currently available for Mac and Linux only. The plugin is available as a VST3 plugin. Simply download the latest release and copy the plugin to your preferred VST3 folder.

In Mac this is:

~/Library/Audio/Plug-Ins/VST3

In Linux this is:

~/.vst3

Building from source

First, clone this repo with its submodules:

git clone --recurse-submodules git@github.com:rodrigodzf/NeuralResonatorVST.git

For all platforms the plugin can be built using:

bash ./bin/build.sh

It is also possible to build the plugin with a traditional C++ GUI using:

bash ./bin/build.sh -s

Dependencies

For building on Linux we need the dependencies of juce:

sudo apt update
sudo apt install libasound2-dev \
    libfreetype6-dev \
    libx11-dev \
    libxcomposite-dev \
    libxcursor-dev \
    libxext-dev \
    libxinerama-dev \
    libxrandr-dev \
    libwebkit2gtk-4.0-dev

Libtorch is automatically downloaded by the build script. However, in case you want to use your own libtorch installation it is possible to manually set it as:

cmake -S . -B build  -DCMAKE_PREFIX_PATH=YOUR_LIBTORCH_PATH

Development

For development, we can build the project using:

sh ./bin/build.sh --test
cd ui && npm run dev

This will open a server on localhost:3000 and you can start developing the UI. The UI will automatically reload when you change something in the source code.

The VST and the UI communicate via a websocket. The websocket server is started by the VST and the UI connects to it. The websocket server is started on localhost:8000/ui.

Logs and config files

The logs are written to the user application data path.

In Mac this is:

tail -F ~/Library/Logs/NeuralResonatorVST/log.txt

and in Linux this is:

tail -F ~/.config/NeuralResonatorVST/log.txt

Related projects

This repo uses extensively the Typescript code from here for the user interface.

For the websocket server it uses the code from here. Because this project does not depend on openssl, it uses instead:

• for sha1 hashing
• for base64 encoding

It also uses the code from here for the pytorch plugin and the filterbank implementation.