fcaspe / BesselsTrick

Really-Real Time FM Tone Transfer Audio Pluigin

Home Page:https://fcaspe.github.io/BesselsTrick/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Logo

Bessel's Trick is a Neural Audio Plugin for fast, live Tone Transformation of Musical Instrument sounds using Frequency Modulation (FM) synthesis.

GitHub release

Features:

  • FM Synth controllable from an Audio Input.
    • Try feedng it voice, violin, solo guitar, trumpet.
  • Designed for live playing.
    • Fast sound generation, 1.3ms inference latency
  • 6-oscillator FM engine.
    • Based on the classic Yamaha DX7.
  • Comes with many pretrained sounds!
    • See the Release webpage.
  • Tweakable FM controls on real time.
  • VST3 and AU plugin formats available.

 

Plugin's GUI

Build it

Clone this repository and run:

source build.sh

This will run the build and install process.

Bringing in more sounds

Bringing more sounds require obtaining an FM patch in DX7 format and then training a new model. For info on how to train new models visit the Envelope Learning repository.

How does it work?

Logo

Bessel's Trick does not use MIDI to control the synth. Instead, it works in an audio to audio fashion, similar to other Tone Transfer approaches.

The plugin extracts continuous sequences of pitch and RMS from a monophonic audio signal, and uses small neural nets (one for each sound) to map these to the envelopes of a six-oscillator FM synth.

But, unlike previous approaches, I use a different method for training the neural nets, that addresses the poor articulation present in previous methods. If you want to learn more check out my Envelope Learning paper.

Why FM?

Frequency Modulation synthesis is a versatile and compact approach for synthesizing a wide range of timbres. Moreover, there is a wide variety of sounds already designed, particularly in the form of Yamaha DX7 Patches.

Envelope Learning learns to control envelopes of synthesizer models. We use Yamaha DX7 patches to train the neural nets.

Why Neural Nets?

Each sound is implemented with a really tiny Recurrent Neural Network (RNN) that has been trained to replicate the oscillators' amplitude envelopes from a continuous control signals. Players have continuous influence on the sound throughout the playing action.

This continuous control enables some cool articulations for free! For instance, on a guitar, Bessel's Trick can support String Muting, Hammer-on, Pull-off, and Natural harmonics.

Try it with other monophonic instruments!

Acknowledgements

I would like to thank these lovely folks for their time!

  • Jeff Miller (GUI design tips)
  • Rodrigo Diaz (Building Pipeline, Apple Notarization)
  • Daniel Walz (PGM library tips)

Licenses

  • Bessel's Trick released under BSD-3.
  • Foleys GUI Magic Library licensed under BSD-3.
  • YIN Pitch Tracker adapted from Pidato, licensed under GNU GPL 3.

Citation

If you find this work useful, please consider citing us:

@article{caspe2023envelopelearning,
    title={{FM Tone Transfer with Envelope Learning}},
    author={Caspe, Franco and McPherson, Andrew and Sandler, Mark},
    journal={Proceedings of Audio Mostly 2023},
    year={2023}
}

About

Really-Real Time FM Tone Transfer Audio Pluigin

https://fcaspe.github.io/BesselsTrick/


Languages

Language:C++ 90.0%Language:CMake 7.9%Language:Shell 1.3%Language:Python 0.8%