DCGAN128

This is an implementation of Deep Convolutional Generative Adversarial Networks modified to create 128*128 sized spectrograms to be able to be converted to WAV files using Analysis & Resynthesis Sound Spectrograph or ARSS for short.

Prerequisites

• install ffmpeg
• install arss

pip install tqdm

Usage

Put wav files (songs of the same genre) in a folder and run

python process.py

Follow the prompts after running, process.py will split the wav files into 2 second segments and convert them into spectrograms of file type bmp using ARSS. Then since the bmp files are greyscale the files will be converted to png with just 1 image channel in a folder called png.

python train.py

will prompt the name of the images data (in this case png) and appropriate batch and epoch size.

After saving the model to a pth file generate.py and subsequently bmp_to_wav.py will generate new spectrograms and wav files

128DCGAN-Spectogram/
    └── data
    └── bmp
    └── png
    └── bmp_generated
    └── wav_generated
    └── bmp_to_wav.py
    └── data.py
    └── generate.py
    └── model.py
    └── process.py
    └── train.py

Generated Spectrograms using Songs

Input

Results