How to Set Up Local Voice Cloning and Text-to-Speech on Windows

Generating a custom voice, or cloning one from a short sample, no longer needs a cloud subscription or a datacentre. Local voice cloning and text-to-speech runs comfortably on a Windows PC with a modest NVIDIA GPU, and the whole pipeline stays on your own machine. The amount of VRAM you have decides how much text you can synthesise in one pass and how close to real-time you can run, but the entry bar is lower than most people expect.

Quick Answer

You can run open-weight text-to-speech and voice cloning locally on Windows with an NVIDIA GPU of 4 to 8GB VRAM. Models like XTTS-v2 clone a voice from a 6 to 30 second clean audio clip and synthesise speech across 17 languages, using roughly 3 to 4GB of VRAM at FP16. The setup is a Python install plus a model download.

What you need before you start

The realistic minimum is an NVIDIA GPU with 4GB of VRAM, though 8GB gives you more headroom for longer passages and faster generation. XTTS-v2, one of the most popular open-weight models, uses about 3 to 4GB of VRAM at inference in half precision, which puts it within reach of mainstream RTX cards rather than only high-end ones. You also need Python 3.8 or newer and a working CUDA setup so the model uses the GPU instead of the much slower CPU path.

For a reference voice, prepare a clean clip of 6 to 30 seconds, single speaker, no background noise, spoken at a normal pace. Quality in equals quality out. If your current card is short on VRAM, the GPU best sellers are the place to find a model that clears the bar with room to spare, and purpose-built AI PCs at Evetech pair that GPU power with the memory these models like.

Step-by-step setup on Windows

1. Install Python and a virtual environment. Use Python 3.8 or later, then create a dedicated virtual environment so the TTS packages do not collide with anything else on your system.
2. Install a CUDA-enabled build. Install the GPU build of PyTorch matching your CUDA version first, so the model runs on the NVIDIA card. Confirm the GPU is detected before going further; a CPU-only install will technically work but generates painfully slowly.
3. Install the TTS package. The current maintained package is coqui-tts, installed with pip. Avoid the older abandoned package of a similar name, which causes most of the install headaches people report.
4. Download the model. Pull the multilingual XTTS-v2 model the first time you run it. It downloads once and caches locally, so subsequent runs start quickly.
5. Run a test synthesis. Feed in a short sentence and your reference clip, and generate a sample. To cut VRAM use by roughly half, switch the model to half precision (FP16), which lets a 4GB card handle longer text.
6. Clone and iterate. Point the model at your reference audio and synthesise. A single worker processes roughly 300 to 600 characters per second depending on the GPU tier, so longer scripts are best split into chunks.

Getting clean output

The biggest quality lever is the reference clip, not the model. A quiet, single-speaker recording at a steady pace produces a natural clone, while a noisy or clipped sample produces artefacts no setting will remove. Keep clips short and clean rather than long and messy.

A note on commercial use

XTTS-v2 ships under the Coqui CPML licence, which restricts commercial deployment without separate terms. If you plan to use generated audio in anything revenue-generating, read the licence first or choose a model with a permissive licence for commercial work. For a strictly personal project this is not a concern.

Frequently Asked Questions

How much VRAM do I need for local voice cloning?

A 4GB NVIDIA GPU is the practical minimum, and 8GB gives smoother, faster generation for longer text. XTTS-v2 uses roughly 3 to 4GB at half precision.

How long does the reference clip need to be?

Between 6 and 30 seconds is ideal. It must be clean, single-speaker audio at a normal speaking pace, since the model copies whatever characteristics the sample contains, including any background noise.

Can it generate speech in other languages?

Yes. XTTS-v2 supports 17 languages and can clone a voice in one language then synthesise speech in another while keeping the speaker's vocal character.

Why is generation slow on my machine?

The most likely cause is the model running on the CPU rather than the GPU. Confirm your CUDA-enabled PyTorch build is installed and the NVIDIA card is detected before troubleshooting anything else.

Is local voice cloning free to use commercially?

Not automatically. XTTS-v2 uses the Coqui CPML licence, which requires separate terms for commercial use. Personal projects are fine, but check the licence before deploying anything that earns revenue.