---
language: en
tags:
- audio
- diffusion
- generative
---

# WindGenerator

DDPM diffusion model trained on mel spectrograms of wind recordings.

- Architecture: UNet2DModel, 2.5M parameters
- Training: 74,000 steps on 1,966 wind clips
- Output: ~10s wind audio via Griffin-Lim reconstruction

## Usage
See [GitHub repository](https://github.com/alpercagan/WindGenerator) for full pipeline and Colab demo.