A live style-transfer effect. Sing, play or speak into it and LiveTransfer re-synthesizes the sound in the voice of a trained neural audio model (RAVE), in real time.
The idea
LiveTransfer does not filter your sound. It encodes the incoming audio into the model's own internal space, lets you nudge that encoding, and decodes it back to audio as the model would have made it. Two models are included, and they are different instruments rather than variations: Dawn chorus, trained on birdsong, and Strengjavera, trained on a magnetic resonator piano.
Controls
- Model picks which model your sound is rebuilt in.
- Amount blends the transformed sound with your original. Partial amounts keep the detail of your performance with the model layered in.
- Z1 to Z8 offset eight dimensions of the model's space, from −3 to +3. At zero you hear the model's own reading of your input, the most faithful transfer. These dimensions have no fixed meaning such as "brightness": they are whatever the model learned, and they differ between models. The first ones change the result the most, the last ones shift its colour. Far from zero the sound becomes artefacted and unstable. Center returns all eight to zero.
- Level sets the output.
The panel shows the model's buffer latency and its reference compute cost, and the loading and ready state of the model.
Getting started
Choose a microphone, or play a clip or recording through it. Start with Faithful Dawn for the cleanest transfer, then Half Blend to keep your own sound underneath, Strengjavera for the second model, and Displaced, Late Colour or Outside to steer further away.
LiveTransfer is mono, and the models add latency, so it suits textures and sound design more than tight live timing.
