How to Run Chatterbox TTS on Windows: Easy and Manual Setup (2026)
Chatterbox is a strong open-source voice cloner, but its official instructions target Linux, and Windows installs often fail on PyTorch and CUDA. This guide shows the fastest way to run Chatterbox on Windows, then the errors that break manual installs, then the manual route for people who want the raw model.
Quick answer
- Install the Voice Creator Pro desktop app from the Microsoft Store (it has a free trial) or the Mac App Store, and pick Chatterbox from the model list.
- Clone a voice from a clean 3 to 10 second clip, type your text and generate. Everything runs locally and offline.
- The app also builds full workflows on top of the models: audiobooks with automatic speaker identification and voice assignment, video dubbing, voice changing and a local API.
- For the manual route you need Python 3.11, a GPU build of PyTorch and the
chatterbox-ttspackage from Resemble AI's GitHub.
Two ways to run Chatterbox on Windows
| Voice Creator Pro desktop app | Install it yourself | |
|---|---|---|
| Models | Chatterbox Multilingual and Turbo | Chatterbox, Multilingual, Turbo |
| Setup | Install the app | Python, PyTorch, CUDA |
| Languages | 23 (Multilingual) | Up to 23 |
| Runs offline | Yes | Yes, once set up |
| Cost | One-time purchase | Free model |
Step 1: Install the Voice Creator Pro desktop app
Voice Creator Pro is a desktop app for Windows and Mac with Chatterbox built in, alongside OmniVoice, Qwen3-TTS, VoxCPM2 and DramaBox. Chatterbox still needs Python, PyTorch and GPU libraries underneath, but the app ships them already matched, so there is nothing to install or debug.
Download it from the Microsoft Store, which has a free trial, or buy it on itch.io. On a Mac, get it from the Mac App Store. It is a one-time purchase starting from $54.99 (checked Oct 2026), with no subscription and unlimited generations.
On Windows the app runs on your graphics card, with 8 GB of video memory recommended, or on the CPU alone, more slowly. On a Mac it needs Apple Silicon (M1 or later).
Step 2: Pick Chatterbox in the model list
Open the Lab and click the model name to switch models. Voice Creator Pro offers two Chatterbox versions:
- Chatterbox Multilingual: the higher-quality version, with conversational speech in 23 languages.
- Chatterbox Turbo: faster, English only, lower quality. Select Chatterbox, then set quality to Lower.
Start with Multilingual unless speed matters more than quality.
Step 3: Clone a voice and generate
Drop in a clean reference clip of 3 to 10 seconds, type or paste your text, choose the language and click generate. Voice Creator Pro clones zero-shot, so there is no training step, and a longer reference clip does not make a better clone. Our guide on how much audio voice cloning needs covers what makes a good sample.
Because the desktop app runs Chatterbox locally and offline, your reference clips and scripts never leave your PC.
Clone your first voice with Chatterbox on your own PC
See desktop pricesStep 4: Tune delivery with exaggeration, pacing and tags
Chatterbox has two controls in the advanced settings (the sliders icon next to the language selector):
- Exaggeration (default 0.50): higher values sound more animated, lower values more neutral.
- CFG Weight (default 0.50): sets the pace. Around 0.2 to 0.3 speeds speech up; 0.7 to 0.8 slows it down.
Resemble AI suggests pairing a higher exaggeration (about 0.7) with a lower CFG weight (about 0.3) for dramatic reads, since higher exaggeration tends to speed speech up. You can also click the smiley icon to insert expression tags, such as a laugh, straight into your text.
More than a model: the workflows built around it
The open-source model turns a short passage of text into speech, and everything else is up to you. Voice Creator Pro builds complete workflows on top of Chatterbox and its other voice models:
- Audiobooks: import an EPUB, PDF, Word document or text file and the app handles the whole book. It identifies the speakers automatically and assigns each one a voice, so dialogue doesn't come out in the narrator's voice. Numbers, dates and currency are rewritten into words before they reach the model, and a pronunciation lexicon fixes names that come out wrong across every generation. Read more about long-form audio.
- Video dubbing: dub a video into 21 languages. Each speaker's voice is cloned from the video automatically, lines are timed to fit the original, and the background music stays under the new dialogue.
- Voice changer: the voice changer uses the Chatterbox engine to re-voice a recording you already have while keeping your timing and emotion.
- Local API: automate narration, cloning and voice design from your own scripts through the desktop app's REST API, offline.
Common problems when installing Chatterbox yourself
Most manual installs fail before Chatterbox ever loads. These are the problems Windows users report most often.
PyTorch installs without GPU support. A plain pip install often pulls the CPU-only build of PyTorch, so Chatterbox ignores your graphics card and a sentence takes minutes. Check whether PyTorch can see CUDA before blaming the model.
Torch and torchvision don't match. An error like torchvision::nms does not exist means the PyTorch packages were built for different versions, usually after a partial upgrade. The fix is to uninstall torch, torchvision and torchaudio together and reinstall them as one matched set.
New RTX 50-series cards aren't supported by the pinned PyTorch. Chatterbox pins its PyTorch version, and older builds don't include support for the newest NVIDIA cards. Upgrading PyTorch can then conflict with the version Chatterbox asks for.
The wrong Python version. Resemble AI tests on Python 3.11. Newer Python versions may lack Windows builds of some libraries, which shows up as failed installs.
"No module named chatterbox". The virtual environment isn't active in the current terminal. Each new PowerShell window needs it activated again.
Model downloads fail. Chatterbox downloads its weights from Hugging Face on the first run, and some setups need a free Hugging Face access token to do it.
Robotic or accented clones. This one looks like an install problem but comes from the reference clip. Background noise and echo get copied into the clone. If the reference language differs from the output language, Resemble AI suggests setting CFG weight to 0 to reduce accent carry-over.
The manual route, in brief
Running Chatterbox yourself makes sense if you want to script it into your own pipeline. We point to the official source for the exact commands, since versions change with each release.
Hardware. An NVIDIA GPU with CUDA. Resemble AI doesn't publish a VRAM figure; community reports put the multilingual model at around 4 to 6 GB. It runs on the CPU too, much more slowly.
The general shape of the process:
- Install 64-bit Python 3.11 and create a separate virtual environment for Chatterbox.
- Install a CUDA build of PyTorch that matches your NVIDIA driver.
- Install the
chatterbox-ttspackage and let it download the model weights on first run. - Run the example script or the included Gradio demo, then clone from your own reference clip.
Where to get the exact steps: the official Chatterbox repository on GitHub, published by Resemble AI. Its README lists the current models and install commands. The project has been released under the MIT license; check the current terms there before commercial use. The README also says every file it generates carries Resemble's Perth watermark.
Who should run Chatterbox themselves
Install it yourself if you are comfortable with Python, virtual environments and CUDA errors, and want to call the model from code.
Use the desktop app if you want a graphical interface on Windows, finished workflows like audiobooks with a voice per speaker or video dubbing, or other models to compare on the same voice. Our Qwen3-TTS vs OmniVoice vs Chatterbox test scores all three on the same cloned voice.