Best Local Text-to-Speech Models in 2026: Sizes, Licences and What Runs Offline

Local text-to-speech has quietly become practical. A usable English voice is now a 63 MB file that runs on a CPU, and the best-sounding open model is 327 MB. The catch is not quality or hardware any more — it is licensing, and one of the most widely used models is non-commercial in a way that is easy to miss, because the licence it points at no longer exists on the web.

Quick answer

• Smallest useful: Piper, about 63 MB per English voice, runs on CPU. Two licences apply: the code is GPL-3.0, the voice files are MIT.
• Best general quality: Kokoro-82M, a single 327 MB file, Apache-2.0, and by far the most downloaded.
• Voice cloning: XTTS-v2 — but it is non-commercial only, and the project behind it is gone.
• Read the licence before the samples. The three leading options are GPL-3.0, Apache-2.0 and non-commercial. They are not interchangeable.

The models, by size and licence

File sizes are read from each model’s published files on Hugging Face, checked 23 September 2026. Repository activity is from the GitHub API on the same date.

Piper: the one that fits anywhere

Piper ships one ONNX file per voice, and the English (US) collection alone has 27 voice files across 21 speakers in three quality tiers. The sizes are genuinely small:

• low — 63 MB (5 files)
• medium — 63 to 79 MB (18 files)
• high — 114 to 137 MB (4 files)

That is small enough to sit alongside a language model without competing for VRAM, which is the usual reason people give up on adding speech to a local setup. It runs on CPU, so it does not compete for the GPU at all.

⚠️ Two licences apply, and the distinction matters. The code repository is OHF-Voice/piper1-gpl under GPL-3.0 — the name is not subtle about it — but the voice files in rhasspy/piper-voices are published under MIT. So the models you ship are permissively licensed; it is the engine around them that carries copyleft obligations. GPL-3.0 is a copyleft licence: fine for internal tools and for GPL-compatible projects, but it carries obligations that Apache-2.0 does not if you distribute software built around it. If you are shipping a product, read it rather than assuming “open source” means “do as you like”. Our model licence checker covers what the common licences permit.

Kokoro: the quality pick, and a nuance

Kokoro-82M is a single 327 MB file (kokoro-v1_0.pth) under Apache-2.0, which is the most permissive of the realistic options. It is also, by a wide margin, the most used: 11.7 million downloads in 30 days on Hugging Face, against roughly 700,000 for the community ONNX build and 7.0 million for XTTS-v2.

The nuance worth stating plainly: its GitHub repository has not been pushed to since 6 August 2025. On the maintenance test that would be a warning sign. Here it reads differently — the model is small, finished and enormously used, with active community builds around it (ONNX, MLX and a Chinese variant all published separately). A quiet repository behind a heavily used artefact is not the same as an abandoned tool, but it does mean you should not expect fixes.

🔴 XTTS-v2: non-commercial, and the licence page is gone

XTTS-v2 is the model people reach for when they want voice cloning from a short sample, and it is still downloaded 7.0 million times a month. It is also the one most likely to cause a problem, for two reasons.

First, it is non-commercial. The model card gives the licence as the Coqui Public Model License. The licence text is unambiguous: “This license allows only non-commercial use of a machine learning model and its outputs.” It goes further than most people assume. Its definition of a non-commercial purpose applies “only so far as you do not receive any direct or indirect payment arising from the use of the model or its output”; it also states that using the model “to train other models for commercial use is not a non-commercial purpose”, and excludes charitable “revenue-generating activity, including projects directly funded by government grants”.

Second, the licence it cites has disappeared. The model card links to coqui.ai/cpml, and the licence file itself prints that same URL at the top. The entire coqui.ai domain now returns 404 — not just that page. The only surviving copy of the terms is the LICENSE.txt bundled in the Hugging Face repository. The Coqui TTS codebase is still on GitHub with 46,000 stars, but its last push was 16 August 2024.

So: usable for hobby projects and internal experimentation, not for anything you charge for — and if you need to show a colleague or a lawyer the terms, download the file, because the canonical URL is dead.

What this costs to run

Very little, which is the point. Piper runs on CPU. Kokoro at 327 MB fits beside anything. Only Orpheus, at roughly 3.78 billion parameters, is in the same memory conversation as a language model — and it is a gated download, so you will need to accept terms on Hugging Face before you can fetch it.

If you are adding speech to an existing local stack, the budget question is what is already resident. Size the whole thing with the VRAM calculator, and remember that the context window of your language model is usually the larger cost — see what a context window costs in VRAM.

Where speech fits in a local setup

Text-to-speech is the output half. The input half is speech recognition, and the two together are what turn a local model into something you talk to rather than type at. A private assistant that never sends audio anywhere is one of the stronger arguments for running locally at all — see self-hosted AI for what that does and does not protect, and building a private RAG chatbot if the goal is answering questions over your own documents.

For the model behind it, the runner comparison covers what to serve it with, and how much RAM and VRAM you need covers the hardware floor.

FAQ

What is the best open source TTS model?

Kokoro-82M for quality under a permissive licence — 327 MB, Apache-2.0, and the most downloaded by a wide margin. Piper if size matters more than naturalness.

Can I use XTTS-v2 commercially?

No. Its licence states it “allows only non-commercial use of a machine learning model and its outputs”, and defines that to exclude any direct or indirect payment arising from the output.

How much disk does a local voice need?

A Piper English voice is 63 MB at low and medium quality, up to 137 MB at high. Kokoro is a single 327 MB file.

Does local TTS need a GPU?

Piper does not — it runs on CPU. The larger models benefit from one, but none of these is demanding next to a language model.

Is Kokoro still maintained?

Its repository has not been pushed since August 2025, but the model is current and heavily used, with active community ONNX and MLX builds. Expect no fixes rather than no future.

Sources

• Hugging Face, checked 23 September 2026: Kokoro-82M and XTTS-v2 — file sizes, licence tags and download counts read from the API and repository files
• Coqui Public Model License 1.0.0, quoted from the LICENSE.txt in the XTTS-v2 repository, because coqui.ai no longer resolves
• OHF-Voice/piper1-gpl — Piper licence and repository activity

Updated 23 September 2026. Sizes and licences were read from the published model files on that date. Licences change and projects move; check the model card before shipping anything.