genre-transfer-pipeline

The genre transfer pipeline system.

Requirements

The system has only been tested on the GNU/Linux operating system. I cannot comment on whether it works on Windows/Mac or not (though I suspect it should—provided you have access to the same command-line utilities).

Requires:

• GNU Bash
• Python 3.6+
• FFmpeg
• Nvidia GPU with CUDA 10+

Installation

All code dependencies can be installed by running the following. It's recommended to create a virtual environment before installing it.

pip install -r requirements.txt

Training

Refer to https://github.com/descriptinc/melgan-neurips for training the spec-reconstruction part of the pipeline and https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix for training the style-transfer system. Instead of mel spectrograms, CQT spectrograms must be used. These can be created using wav-to-spec/wav-to-cqt.py.

Running

Running inference on a single 22.05kHZ piece of music can be done.

./pre-process.sh #Processes a single file to make it ready for the system

After pre-processing the system can be run provided there are relevant trained and saved models set up for the genre you need.

./run.sh [genre1] [genre2] #Runs the entire system, giving you the ability to perform genre transfer.

Licencing

This project incorporates the MIT License for the code located in the spec-reconstruction directory and the BSD License for code located in the style-transfer directory. All other code is licenced via the MIT license.