- Three licences hide inside every voice project
- What you're self-hosting, grouped three ways
- Eight engines and models compared
- The eight, in no particular order
- Will it run offline, on a CPU, or only on a GPU?
- My advice on picking a voice to run yourself
- When a hosted voice is the saner route
- Frequently asked questions
Open source text to speech means a voice engine or model you download and run on your own computer or server, with no hosted service in between. Before you use one for business audio, check three licences, not one: the code, the model weights and the voice data, because a permissive code licence doesn't always cover the voice itself.
Choosing an open source voice comes down to three decisions: what you're allowed to do with the audio, what hardware you'll run it on, and how much setup you'll put up with. Get those right and the shortlist is small. Get them wrong and you'll end up with a lovely voice you can't use for client work, or a model your laptop can't run.
Below are eight engines and models, grouped by what you're self-hosting. They're in no particular order, and none of them is a big brand, because the open source rule removes the hosted voice companies. Expect Python, Docker or a decent graphics card for most of them, and know that the lightest options sound the most robotic.
Three licences hide inside every voice project
On a text to speech project, open source can mean three different things. The engine code has a licence. The trained model weights have a licence. And the recordings the model learned from have terms of their own, which often shape the licence on the weights. A project can be open source at the code level and still forbid commercial use of its voices.
One cloning project here shows the split clearly: its README says the code is MIT, but the pre-trained models are CC-BY-NC because of their training data. Another toolkit's code is MPL 2.0 while each pretrained model carries its own terms. And the GPL engines raise a different question, which is what happens if you ship them inside a closed product. So read the licence on the exact model or voice file you'll use, not just the headline on the repository.
What you're self-hosting, grouped three ways
Complete engines that run on a laptop, no GPU needed
Licence line: GPL and Apache. Piper is a fast local neural engine under GPL-3.0; its README also says the Open Home Foundation is looking for new maintainers, so check recent activity before you build on it. eSpeak NG is GPL 3 or later, compact, multilingual and robotic by design. Sherpa-ONNX is Apache 2.0 and runs speech models offline on-device, including in the browser.
Small models and toolkits you load with Python
Licence line: Apache weights and MPL code. Kokoro is a small open-weight model whose card says its Apache weights can be deployed commercially; its repo was last updated in Aug 2025. The Idiap fork of Coqui TTS is the maintained branch of the original toolkit, whose own repository hasn't been updated since Aug 2024; the fork's code is MPL 2.0, but each pretrained model has its own terms.
Cloning and dialogue research models for heavier hardware
Licence line: MIT, Apache and one open core split. F5-TTS ships MIT code with non-commercial pre-trained models. Dia is Apache 2.0 and its README says it's been tested only on CUDA GPUs, with CPU support listed as coming. OpenVoice is MIT for its first two versions, which the README says are free for commercial use; its repo was last updated in Apr 2025.
Eight engines and models compared
| Tool | What you run | Good for |
|---|---|---|
| Coqui TTS (Idiap fork) | Python toolkit, MPL-2.0 code | Pretrained voices, XTTS cloning and very wide language coverage |
| Dia | Dialogue model, Apache 2.0 | Multi-speaker dialogue generated in a single pass |
| eSpeak NG | Engine, GPL 3 or later | Tiny footprint and more than 100 languages, robotic sound |
| F5-TTS | Cloning model, open core | Voice cloning research with MIT code and non-commercial models |
| Kokoro | Small model, Apache 2.0 | A small model its card says can go in commercial projects |
| OpenVoice | Cloning model, MIT | Copying a reference speaker's tone across languages and accents |
| Piper | Local engine, GPL-3.0 | Fast local neural voices from the command line or Python |
| Sherpa-ONNX | Offline toolkit, Apache 2.0 | On-device speech on Android, desktop and in the browser |
The eight, in no particular order
Coqui TTS (Idiap fork)
Best for: Python developers who want a maintained toolkit with pretrained voices, XTTS cloning and models covering around 1,100 languages, after checking each model's own terms.
Dia
Best for: Creators generating multi-speaker dialogue, such as podcast-style exchanges, in one pass, on a machine with a CUDA GPU, the only setup its README says it's tested on.
eSpeak NG
Best for: Accessibility tools, embedded devices and prototypes that need speech in more than 100 languages from a tiny program and can live with a robotic voice.
F5-TTS
Best for: Researchers and hobbyists exploring flow-matching voice cloning, where MIT code is fine but non-commercial pre-trained models rule out client work.
Kokoro
Best for: Builders wanting a small, light open-weight voice model of 82 million parameters, whose model card says the Apache weights can be deployed in commercial projects.
OpenVoice
Best for: Voice cloning experiments that copy a reference speaker's tone across several languages and accents, under an MIT licence the README says covers commercial use.
Piper
Best for: Home automation, screen readers and local assistants needing a fast neural voice you install with pip and run from the command line, a web server or Python.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Sherpa-ONNX
Best for: App developers shipping offline, on-device speech to Android, desktop or the browser, with bindings for many programming languages and no network needed.
Will it run offline, on a CPU, or only on a GPU?
- Offline: everything here runs on your own machine once the code and model files are downloaded, so no text leaves your computer.
- CPU only: the complete engines are built for ordinary laptops, small servers and phones. The toolkit fork also offers a Docker image for CPU.
- GPU: the dialogue model lists VRAM needs by precision and has only been tested on CUDA. Treat the cloning research models as heavier work too, and check each README before you buy hardware.
- Setup: expect a terminal. Most of these install with pip or Docker, and none has a polished desktop app of its own.
My advice on picking a voice to run yourself
If you need speech for an app, a screen reader or a home assistant, start with a complete engine that runs on a CPU, and only move up to heavier models if the voice quality isn't good enough. If you want natural narration for business audio, pick a model whose weights, not just its code, are under a licence that allows commercial use, and keep a note of which file you used and where its licence came from.
If you're cloning a voice, get written permission from the person whose voice you're copying, whatever the licence says. And check how active a repository is before you depend on it: two projects here haven't been updated in over a year, and one is asking for new maintainers.
When a hosted voice is the saner route
Self-hosting suits people who want control, privacy or offline speech, and who don't mind the setup. If you need polished voices this afternoon without touching a terminal, a hosted service will get you there faster. Our guide to hosted AI voice generators covers that side, and if you only need the odd clip, these free ways to turn text into speech online skip the install altogether.
Do you make an open source voice engine or text to speech model? Tool makers can ask to be considered for a featured place on this list.
Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.
Frequently asked questions
Can I use open source text to speech commercially?
It depends on the licence of both the code and the model you use. The Apache and MIT projects here say commercial use is allowed, one cloning project's pre-trained models are non-commercial, and toolkit models can each carry their own terms. GPL engines are fine to use, but shipping them inside a closed product brings obligations.
Does open source text to speech work offline?
Yes. Every engine and model on this page runs on your own machine once you've downloaded the code and voice files, so you don't need an internet connection to generate speech.
Do I need a GPU for open source text to speech?
Not for the lighter engines, which run on ordinary laptops and even phones. The dialogue model here has only been tested on CUDA GPUs, and voice cloning models generally want stronger hardware.
Why do some open source voices sound robotic?
Older formant synthesis builds speech from rules rather than recordings, which keeps it tiny and multilingual but robotic. Neural models sound more natural but need more memory and setup.
Is it legal to clone someone's voice with an open source model?
The software licence only covers the software. Using a real person's voice raises consent and personality rights questions that vary by country, so get written permission and take advice for anything commercial.
Why does a model card warn about websites using its name?
Popular open models attract copycat sites that borrow the name. One model card on this page warns that some sites using its name are likely scams, so download from the official repository or model page only.
