Training recipe for fine-tuning Kokoro-82M for German with a patched StyleTTS2 submodule.

• A reproducible fine-tuning workflow (dataset prep -> Stage 1 -> Stage 2 -> voicepack extraction)
• Original scripts for data preparation and checkpoint/voicepack conversion
• A patched StyleTTS2/ submodule with the fixes required for stable Stage 2 training

• Not a general-purpose Kokoro replacement repository
• Not a bundled upstream mirror of demo/, examples/, kokoro.js/, or tests/
• Not a redistributable training dataset

Start with docs/TRAINING_GUIDE.md.

Go to docs/TROUBLESHOOTING.md.

See docs/ARCHITECTURE.md.

The end-to-end pipeline is working:

Dataset preparation -> Weight conversion -> Stage 1 -> Stage 2 -> Voicepack extraction -> KModel inference

dida-80b/kokoro-deutsch-hui-base is available on HuggingFace.

This is a base model, not a finished single-speaker voice.

# Ubuntu/Debian
sudo apt-get install espeak-ng libsndfile1

# macOS
brew install espeak-ng libsndfile

git clone --recurse-submodules https://github.com/semidark/kokoro-deutsch
cd kokoro-deutsch
uv sync

kokoro/              # Kokoro fork submodule (contains the `kokoro/` Python package)
StyleTTS2/           # Patched training code (git submodule: semidark/StyleTTS2)
scripts/             # Dataset prep, voicepack extraction, inference testing
configs/             # Training config(s)
docs/                # Training guide, troubleshooting, architecture notes
training/            # Local training artifacts metadata (audio excluded)

Contributions are welcome, especially:

• Reproducible runs on public datasets
• Fine-tuning recipes for other languages
• Training stability and quality improvements

See NOTICE for upstream attribution and license details.

Apache License 2.0 — see LICENSE.