Docs / Installation

Installation

Prepare Python, FFmpeg, PyTorch, and your compute device.

Requirements

  • Windows 10 or 11, 64-bit
  • Python 3.11–3.14, 64-bit
  • FFmpeg Shared and FFprobe
  • Approximately 4 GB of free storage
  • NVIDIA GPU recommended; CPU is supported but significantly slower

Install FFmpeg Shared

Open PowerShell and install the shared FFmpeg build required by TorchCodec:

winget install --id Gyan.FFmpeg.Shared -e

Reopen PowerShell after the installation. The static Gyan.FFmpeg package is not sufficient because it does not provide the shared DLLs TorchCodec requires.

Install LLMVoice

Download the latest official release installer from the canonical website endpoint, then run the downloaded file separately:

irm https://llmvoice.yusufdere.com/install.ps1 -OutFile install.ps1
powershell -NoProfile -ExecutionPolicy Bypass -File .\install.ps1

What the installer does

The installer:

  • verifies Windows x64 and detects Python 3.11–3.14;
  • detects supported NVIDIA hardware and chooses a CUDA or CPU runtime profile;
  • downloads the release manifest and SHA256-verified application wheel;
  • creates a versioned isolated runtime under %LOCALAPPDATA%\LLMVoice\runtime;
  • validates FFmpeg shared DLL bootstrap, package consistency, runtime imports, CUDA when selected, and synthetic audio decoding;
  • publishes the launcher and install state atomically;
  • adds only %LOCALAPPDATA%\LLMVoice\bin to User PATH;
  • preserves existing voices, models, cache, and configuration.

It does not modify machine PATH, install into global Python, or install the CUDA Toolkit.

Installer options

Automatic runtime selection is the normal choice. Advanced options can pin a profile or release:

.\install.ps1 -Runtime auto
.\install.ps1 -Runtime cuda
.\install.ps1 -Runtime cpu
.\install.ps1 -Version 0.1.7
.\install.ps1 -Force
.\install.ps1 -Debug

Use -Runtime cuda only when the installer-managed validation can confirm the selected CUDA profile. CPU remains available for supported systems but is slower for long-form generation.

Validate installation

Open a new terminal so the User PATH update is visible:

llmvoice --version
llmvoice doctor

The installer itself needs network access for release assets. The first synthesis may also download XTTS-v2 model weights and request acceptance of the model’s CPML terms.

Development installation

Editable source installs are intended for contributors, not normal users. See the LLMVoice repository README for the development environment and validation workflow.