Danand / audio-sample-generator

Generating unique one-shot audio samples with Stable Diffusion.

Audio Sample Generator

This app can generate new audio samples using the Stable Diffusion model (LoRA) for training.

Steps:

1. Convert existing audio samples to spectrograms.
2. Train a small Stable Diffusion model (LoRA) on spectrograms.
3. Generate new audio samples with a specified prompt.

Prerequisites

• Python

How to install

git clone --recurse-submodules [e-mail]:Danand/audio-sample-generator.git
cd audio-sample-generator
chmod +x run.sh

How to launch

./run.sh

How to use

Simply follow all pages from the sidebar sequentially.

Advanced settings are skipped here for convenience.

Extract Spectrograms

1. Open audio files.
2. Click the Extract button.
3. Review the spectrograms extracted from the audio files.
4. Proceed to the next page.

Prepare Dataset

1. Specify for each spectrogram:
   • Subject
   • Caption (comma-separated keywords)
   • Optional: Weight
2. Click the Save button.

Train LoRA

1. Click the Train button.

Generate Audio with Stable Diffusion

1. Type in the Prompt.
2. Specify the Amount of audio to generate.
3. Click Generate.
4. Listen and save the generated samples if desired.

Extras

Batch Convert to Audio

That page is convenient for batch converting spectrograms to audio samples. You can experiment with any images of the respective size, not necessarily spectrograms.