Conditional audio synthesis using adversarial networks

In this project, we adapt the GANSynth model to work with a drums dataset. We also propose a variant to condition the training in order to obtain a model capable of generating audio based on the desired genre.

Install

To run our scripts, you will need to have a specific environment which will require the installation of miniconda (or anaconda). If you do not already have it, you can install it from the original website.

• Clone the github repository

git clone https://github.com/PierreChouteau/Conditional-audio-synthesis-using-adversarial-networks

• Create a virtual env with python 3.10.8:

conda create -n gansynth python=3.10.8

• Activate the environment:

conda activate gansynth

• Go into the repository and install the dependencies:

cd Conditional-audio-synthesis-using-adversarial-networks
pip install -r requirements.txt

Training

To start a default training (with the default configuration), simply run the script train.py:

python train.py

If you want to modify the configuration (model_name, optimizers, etc...), you will have to change the config file (default_config.yaml ), with the parameters you want.

Model Generation

You can find some of our sounds generated by our GANSynth here.

Project Structure

Conditional-audio-synthesis-using-adversarial-networks
├── configs
│   └── __init__.py
│   └── config.py # global settings for our model
├── data   
├── datasets   
│   └── __init__.py
│   └── dataset.py # script to load datasets defined as Class object
│   └── helper.py # helper script for plot
├── generated_audio # folder containing some of our audio results
├── models  # folder containing models architecture defined as Class object
│   └── __init__.py
│   └── discriminator.py # discriminator architecture
│   └── gansynth.py # script containing the training of the model
│   └── generator.py # generator architecture
├── notebooks 
├── runs # directory to store the data from tensorboard
├── trained_model # directory to store the models checkpoints
├── .gitignore
├── default_config.yaml # default_configuration for training our gansynth model
├── README.md
├── requirements.txt # python dependencies necessary for the project
└── train.py # script to run the training

Hydra color

Configuration files are written in YAML and use the Hydra framework. If there is a problem with the color of the log in the terminal, you can fix it by installing the following package:

pip install 'git+https://github.com/AlsonYang/hydra/@fix/colorlog_filehandler#egg=hydra-colorlog&subdirectory=plugins/hydra_colorlog'