F5-TTS voice cloning: free to run, but the official model is not for commercial use

By Inam. I make VoiceClone, a Windows voice cloning app, so the last column is my own product and I am biased about it. I keep my other engineering notes on designesh.com. Last updated 13 October 2026.

If you have searched for a free voice cloner lately, F5-TTS is one of the names that keeps coming up, and the demos really are good. You give it a few seconds of someone talking and it reads any text back in that voice. What most guides skip is the line at the bottom of its own README about the licence.

F5-TTS is free to play with, but if you want to earn money with the voice it makes, the official model is not allowed. The code is MIT, which is a very open licence, but the trained model you actually download is under CC BY-NC, and NC means non commercial. So for a monetized YouTube channel, a client voiceover or anything you sell, you need a different model.

What is F5-TTS?

F5-TTS is a text to speech model published on GitHub by a group of researchers, with the X-LANCE lab at Shanghai Jiao Tong University among them. Text to speech means you type a sentence and the computer says it out loud. F5-TTS also does voice cloning, which means it says your text in a voice copied from a recording you give it.

It came out in October 2024 with a research paper, and a better v1 base model followed on 12 March 2025. Under the hood it uses a method called flow matching, which you can think of as starting from noise and shaping it step by step into speech, a bit like an image generator does with pictures. The project is popular: when I checked on 13 October 2026, it had about 15,400 stars on GitHub.

Can you use F5-TTS for commercial work?

Not with the official model. The README says it plainly at the bottom: the code is released under the MIT licence, and the pre-trained models are licensed under CC BY-NC because of the training data, a large collection of real world speech called Emilia. The model page on Hugging Face carries the same cc-by-nc-4.0 tag.

This is where people get confused, so here is the difference in normal words. The code is the program, and the model is the trained brain the program loads. You are free to use and change the program for anything, but the brain is what makes the voice, and CC BY-NC 4.0 says you may not use it mainly for commercial advantage or payment.

Two examples make it clear. You test F5-TTS on your own voice at home to see how good cloning has become: that is fine. You use it to narrate a monetized YouTube video, a paid audiobook or a client's ad: do not use the official F5-TTS model for that. And fine tuning does not wash the licence off, because a fine tune starts from those same weights. If real money depends on the answer, ask a lawyer and not a blog post, including this one.

How does F5-TTS clone a voice from a short clip?

F5-TTS works zero shot, which is the technical way of saying it does not train on your voice first. You give it a reference clip and the words spoken in that clip, then the text you want, and it answers in that voice straight away. If you leave the reference text empty, it runs a speech recognizer to write it down for you, which uses extra graphics card memory.

The project's own inference guide has a few rules worth knowing before you start. Keep the reference clip under 12 seconds, because anything longer gets cut to about 12 seconds, and leave about 1 second of silence at the end so it does not chop a word in half. One generation can only be about 30 seconds long in total, counting your clip, but the command line tool and the web app split longer text into chunks for you.

There are also two small habits that save you a lot of retries. Capital letters get read out one by one, which is right for something like K.F.C. but wrong for a normal word, so type ordinary words in lowercase. And a comma or a full stop followed by a space is how you give it a pause. If your result comes out as pure silence, check that FFmpeg is installed, which is the fix the guide itself gives.

How hard is F5-TTS to install on Windows?

It is a normal Python project, so it is easy if you already work in Python and slow if you do not. The README asks you to make a separate environment with Python 3.10 or newer, install FFmpeg, then install the PyTorch build that matches your graphics card, and only then run pip install f5-tts. After that, f5-tts_infer-gradio opens a small web page on your own PC where you upload the clip and type your text.

The step where Windows users usually lose an evening is PyTorch. It has to match your card and your CUDA version, which is NVIDIA's software layer that lets the card do AI work, and a mismatch gives errors that do not explain themselves. There is also a Docker image, but on Windows that is one more program to learn. F5-TTS runs on NVIDIA, AMD, Intel and Apple chips according to the README, and AMD support is only on Linux.

Is there a version of F5-TTS you can use commercially?

Yes, there is one, with a catch. OpenF5-TTS Base is a community model trained with the same F5-TTS code, but on permissively licensed data, so it is released under Apache 2.0, and that allows commercial and personal use. You load it into the same F5-TTS program.

The catch is in its own model card. It is labelled as still in alpha, its author says it is still weaker than the official non commercial model, and it was trained on English speech only. So you can legally sell with it today, but you may not like how it sounds yet. Its author describes it as a base for further fine tuning rather than a finished voice.

F5-TTS or GPT-SoVITS: which one should you use?

Both clone from a short clip, both are free to download, and both need some Python patience on Windows. The real difference for most people is the licence. GPT-SoVITS is under the MIT licence, and its strength is fine tuning: you give it about 1 minute of one voice, it trains on it, and the result keeps much more of the real accent and rhythm than any instant clone.

So this is how I would choose. For testing or for fun, try F5-TTS. If the voice will ever earn money, use GPT-SoVITS or another model whose licence allows it. If you want instant cloning with a commercial licence and do not mind a hidden watermark, I wrote about Chatterbox TTS, which is MIT. And I explain how GPT-SoVITS works in GPT-SoVITS in plain words.

To be fair to F5-TTS, it does some things better. The speech flows very naturally from a tiny clip, it handles English and Chinese in the same model, and its web app can do several voices and styles in one script. For pure research or a personal test, it is one of the best free options there is.

Where does VoiceClone fit?

VoiceClone is the Windows 10 and 11 app I built, and it runs GPT-SoVITS fully offline, so nothing you record is ever uploaded. You double click START VoiceClone.bat and it sets itself up, with no Python for you to install. You record 1 to 3 minutes of your voice, the app checks the recording, and then Make it perfect trains the voice on your own PC, faster with an NVIDIA card and still working on a normal processor.

It is English only and it does not sing. The free tier lets you clone and fully train one voice, then make 3 generations to check it. If it sounds like you, the $49 one time license opens unlimited length, the word editor, retraining and exports, with no subscription, and what you make with your own voice you can use commercially. Please only clone your own voice, or a voice whose owner has clearly agreed. The app is on the download page.

FAQ

Is F5-TTS free for commercial use?

No, not with the official model. The F5-TTS code is under the MIT licence, but the pretrained model is under CC BY-NC 4.0 because of its Emilia training data, and that licence does not allow commercial use. The community OpenF5-TTS Base model is under Apache 2.0 and does allow it.

How long should the F5-TTS reference audio be?

Under 12 seconds, with about 1 second of silence at the end. Longer clips are cut to roughly 12 seconds, which can chop a word in the middle and hurt the result.

Does F5-TTS need a GPU?

It runs best on a graphics card, and the README covers NVIDIA, AMD on Linux, Intel and Apple chips. You install the PyTorch build that matches your device before installing F5-TTS. Without a supported card it is much slower.

Is F5-TTS better than GPT-SoVITS?

For a quick clone from a few seconds, F5-TTS sounds very natural. For one voice you will use again and again, fine tuned GPT-SoVITS usually keeps more of the real accent, and its MIT licence lets you use the result commercially.

Sources: the F5-TTS README and inference guide on GitHub and the F5-TTS model card on Hugging Face, all read on 13 October 2026, the OpenF5-TTS Base model card on Hugging Face, the Creative Commons BY-NC 4.0 licence deed, the GPT-SoVITS LICENSE file on GitHub, and the VoiceClone product page on myvoiceclones.com.