Overview

Built for PiperTTS (custom model training) with a simple file structure:

dataset.zip/
├── metadata.csv
└── wavs/
    ├── <name>_processed<index>.wav
    └── ...

Works with LJSpeech-compatible TTS engines too!
metadata.csv looks like:

wav_filename|transcript
wavs/<name>_processed<index>.wav|<transcript>
...

---

Setup

Tested on Ubuntu Server 22.04 (Python 3.10.12) and Ubuntu Desktop 24.04 (Python 3.12.3). Should run fine on Debian-based systems.
No (official) Windows support, sorry!

1. Clone it:

   git clone https://github.com/DominicTWHV/LJSpeech_Dataset_Generator.git

2. Setup:

   cd LJSpeech_Dataset_Generator
   chmod +x pipeline.sh

3. Run it:

   ./pipeline.sh

   Then hop onto the Gradio WebUI @ port 7860. The server listens on 0.0.0.0:7860 by default. For local use, connect at https://127.0.0.1:7860/ .

---

Post-Processing

Move dataset.zip to your training directory:

mv /output/dataset.zip /path/to/training/dir
unzip dataset.zip

Or just download it via the WebUI.

---

Troubleshooting

File permission issues? Missing files? Check script permissions or background processes.
Still stuck? Feel free to drop a note in the issues tab.