Installing Engines

Most engines are ready the moment their weights download. Four — Chatterbox, OmniVoice, VoxCPM, and Qwen3-TTS — need their own virtual environment because their transformers / torch versions are mutually incompatible (and incompatible with the main app). Voice Studio builds those on demand, so nothing heavy is installed until you ask for it.

Two steps: install, then download

For an isolated engine, the engine menu walks you through both:

1. Install — builds the engine's virtual environment as a subprocess. A dialog streams the live build log (it installs the package, then a CUDA-matched PyTorch build). This is a one-time step per engine.
2. Download — fetches the model weights into the shared Hugging Face cache, with a live progress bar.

In-process engines (VibeVoice, Kokoro, Kitten TTS Mini) skip step 1 — they only need Download.

Removing engines

From the same menu you can reclaim space:

• Delete weights — removes the model files (re-downloadable later).
• Uninstall — removes the engine's whole virtual environment.

Notes

• Environments are built with uv and hardlink shared packages from a repo-local cache, so common builds (like a shared PyTorch) are stored once — total venv disk is roughly halved.
• VoxCPM requires Python 3.10–3.12; the installer checks this before building.
• You can also build an environment from the command line — python studio.py install-voxcpm (and install-chatterbox / install-omnivoice / install-qwen). See the CLI Reference.

For sizes and capabilities, see Engines Overview.