StyleTTS2 (Kokoro-82M Fine-tuning Fork)

Note: This fork's main branch is a patched version of the original yl4579/StyleTTS2 repository. It is maintained specifically as a git submodule for the kokoro-deutsch training recipe.

Why this fork exists

The Kokoro-82M TTS model is based on the StyleTTS 2 architecture. However, to fine-tune it using its published HuggingFace weights, several modifications to the upstream training code are required:

1. PyTorch API Migration (Critical): Migrated torch.nn.utils.weight_norm and spectral_norm to the modern torch.nn.utils.parametrizations API. This is mandatory for compatibility with Kokoro's inference pipeline (KModel).
2. Kokoro Symbols: Integrated Kokoro's specific 178-token IPA vocabulary (kokoro_symbols.py).
3. Bug Fixes:
   • Fixed an unsqueeze shape mismatch crash at epoch boundaries involving F0 tensors.
   • Fixed checkpoint saving order to prevent data loss if TensorBoard audio generation fails.
   • Fixed missing .train() mode re-initializations after checkpoint loading in Stage 2.
   • Removed hardcoded ipdb breakpoints that caused silent hangs.
   • Added a monkey-patch for torch.load weights_only=False for PyTorch 2.6+ compatibility.
   • Filtered long phoneme sequences (> 510 tokens) to prevent PLBERT position embedding overflows.

Usage

This repository is not meant to be used standalone.

Please see the kokoro-deutsch repository for the full end-to-end training guide, dataset preparation scripts, and voicepack extraction tools.

---

For the original StyleTTS2 project and documentation, please visit the upstream repository.