CS570 Artificial Intelligence and Machine Learning

Repository for Group Project of CS570, KAIST, 2023 Spring

Collaborator

Sorted by Korean alphabetical order

• Hyeongseok Gwak
• Seongho Keum
• Suhwan Sung
• Sungwoo Jeon

About the Course

Please refer to the syllabus

About the Project

Diffusion-GAN Model for Audio Synthesis

Report

Overview of architecture

Demo

Our Diffusion-GAN melspectrogram result (100 samples) is in diffgan_output.zip. You can convert each melspectrogram .png file into .wav

About the Code

Pre-processing

Use wav2mel.py to convert from audio.wav into mel.png

$ python wav2mel.py --input audio.wav --save mel.png

Training

We utilized the publicly accessible source code of Diffwave and Diffusion-GAN for our project

Our experimental setup is described in our report.

Convert Mel-spectrogram to audio

Use mel2wav.py to convert from mel.png into output.wav

$ python mel2wav.py --input mel.png --save output.wav