Requirements

Agent Avatar does all of its work — voice training, speech generation, and video rendering — on your own machine. That means the requirements are real, and worth checking before you start.

GPU

An NVIDIA GPU with 8GB or more of VRAM. Everything that matters runs on CUDA: voice fine-tuning, F5-TTS inference, and MuseTalk lip sync. More VRAM lets you raise the batch size when rendering a long script.

AMD and Intel GPUs are not supported. There is no CPU-only mode that is worth using.

System

• 32GB of system RAM.
• Roughly 100GB of free disk, most of it for model weights. Your own recordings, trained checkpoints, and rendered clips add to that over time.
• A modern multi-core CPU. Audio preprocessing and video muxing are CPU-bound.

Operating system

Linux, or Windows with WSL2. Agent Avatar is developed and tested on WSL2 with Ubuntu.

There is no native Windows build and no macOS build. On Windows you run it inside WSL2, which needs a working CUDA passthrough to your NVIDIA GPU — that is supported by NVIDIA's WSL driver, but it is a setup step, not automatic.

Software

• Conda (Miniconda is fine). The install creates three separate environments, because the voice, lip sync, and photo animation engines need incompatible dependency sets.
• Python 3.10, installed into those environments by the setup script.
• ffmpeg, for audio and video handling.
• git, to fetch the third-party engine repositories.

What you supply

• 15–30 minutes of clean audio of yourself speaking. This is the single biggest factor in how good your cloned voice sounds.
• A short piece of front-facing video of yourself, for MuseTalk lip sync — or a single well-lit, front-facing photograph, for SadTalker.

Checking your machine

Once installed, ./avatar doctor verifies the environments, models, and GPU access, and tells you what is missing.