Neural Modeling of Magnetic Tape Recorders

This repository contains the code for the research article:

O. Mikkonen, A. Wright, E. Moliner and V. Välimäki, “Neural Modeling Of Magnetic Tape Recorders,”
in Proceedings of the International Conference on Digital Audio Effects (DAFx),
Copenhagen, Denmark, 4-7 September 2023.

The pre-print of the article can be found from here.
The accompanying web page can be found from here.
The datasets can be found from here.

• NEURAL MODELING OF MAGNETIC TAPE RECORDERS
  • SETUP
  • USAGE
  • CITATION
  • ACKNOWLEDGMENTS

SETUP

Clone the repository and submodules

git clone git@github.com:01tot10/neural-tape-modeling.git
cd neural-tape-modeling
git submodule init && git submodule update

Create the Python virtual environment with mamba (preferred) / conda

mamba/conda env create --file environment.yaml

Download data to symlinked location audio/ -> ../neural-tape-audio/

# create a directory for contents
mkdir ../neural-tape-audio
# download and extract toy data
wget -P ../neural-tape-audio 'https://zenodo.org/record/8026272/files/neural-tape-audio_CHOWTAPE.tar'
tar -xzvf ../neural-tape-audio/neural-tape-audio_CHOWTAPE.tar -C ../neural-tape-audio/
# download and extract real data
wget -P ../neural-tape-audio 'https://zenodo.org/record/8026272/files/neural-tape-audio_AKAI.tar'
tar -xzvf ../neural-tape-audio/neural-tape-audio_AKAI.tar -C ../neural-tape-audio/

Optional: To generate target audio with CHOWTape, a VST instance of the plugin should be compiled. Check instructions in the corresponding repository.

USAGE

The folder scripts/ contains the various processing pipelines for interacting with the system, as well as a separate README.md with instructions.

CITATION

Cite the work as follows

@conference{mikkonen_neural_2023,
  title = {Neural Modeling of Magnetic Tape Recorders},
  booktitle = {Proceedings of the {{International Conference}} on {{Digital Audio Effects}} ({{DAFx}})},
  author = {Mikkonen, Otto and Wright, Alec and Moliner, Eloi and V{\"a}lim{\"a}ki, Vesa},
  year = {2023},
  month = sep,
  address = {{Copenhagen, Denmark}}
}

ACKNOWLEDGMENTS

🖤🖤🖤

• VST instance of a reel-to-reel tape machine: CHOWTape
• Python VST wrapper: pedalboard
• Dataloader extended and customized from microtcn
• Error-to-signal ratio (ESR) loss from Automated-GuitarAmpModelling
• ESR loss with DC blocker from GreyBoxDRC
• Multi-resolution short-time Fourier transform (STFT) loss from auraloss-library
• Codebase kept clean with with yapf, isort, pylint and beautysh

🖤🖤🖤