---

name: style-tts2 description: "Use StyleTTS2 source-checkout workflows for TTS data/config preparation, staged CUDA training, fine-tuning, pretrained inference, and troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: MIT

StyleTTS2 repo skill

Use this operating skill when a task involves the StyleTTS2 research repository: text-to-speech training, fine-tuning, pretrained LJSpeech/LibriTTS inference, source-checkout dependency checks, or debugging data/config/checkpoint failures.

First decision

• Preparing data lists, OOD text, YAML configs, 24 kHz audio roots, or asset paths: read sub-skills/data-and-config/SKILL.md.
• Launching first-stage training, second-stage training, fine-tuning, or the one-GPU Accelerate fine-tune path: read sub-skills/training/SKILL.md.
• Running or troubleshooting pretrained LJSpeech/LibriTTS demos, phonemizer/espeak setup, reference audio, diffusion/style controls, or voice-use cautions: read sub-skills/inference/SKILL.md.

If the request is unclear, start with data/config validation before attempting a training or inference run.

Repository shape and runtime assumptions

StyleTTS2 is source-checkout research code, not an installable Python distribution. Do not claim that pip install -e . works unless packaging metadata has been added after this skill was generated. Use the checkout as the working tree and run the repository launchers through bundled helpers.

Read references/runtime-overview.md for the distilled repository layout, dependency set, backend expectations, and verified helper APIs. Read references/repo-provenance.md before deciding whether this skill is stale for a different checkout.

Quick runtime check

Before a task that depends on imports or CUDA, run the bundled safe checker from this skill directory or with an explicit checkout path:

python scripts/check_runtime.py --repo-root /path/to/StyleTTS2 --check-cuda

The checker imports source modules, reports missing hidden dependencies, and can optionally load helper assets without starting training, downloading models, or synthesizing audio.

Installation notes

The repository README documents:

pip install -r requirements.txt

For local training/inference, also account for runtime imports and demo dependencies discovered from source inspection:

python -m pip install pandas tensorboard
python -m pip install phonemizer  # inference demos only; also provide espeak-ng or espeak

Use a CUDA-enabled PyTorch/Torchaudio build for training and fine-tuning. Inference can run on CPU or CUDA, but CPU is slower.

Common routes

Task signal	Route	First safe action
list row, root_path, OOD text, config YAML, missing data	data-and-config	Run validate_data_lists.py or inspect_config.py.
train_first.py, train_second.py, train_finetune.py, checkpoint resume, OOM, NaN	training	Run build_training_command.py in dry-run mode.
LJSpeech/LibriTTS demo, reference audio, phonemizer, alpha, beta, diffusion_steps	inference	Run check_inference_assets.py.
import failure, CUDA unavailable, missing pandas/tensorboard, source checkout not importable	references/troubleshooting.md	Run scripts/check_runtime.py.

Safety and ethics

Training jobs are long-running CUDA jobs that write checkpoints/logs and may download WavLM through Transformers. Do not start them without explicit user approval. Pretrained voice outputs must be disclosed as synthesized unless the user has voice permission and license rights.

Router metadata

This skill routes under the speech/audio scenario via references/repo-routing-metadata.json. Do not hand-edit the live repo-skills-router; use the verifier/import helper if the user later asks to import.