What is the best free text-to-speech model?

It depends on the job. For fast offline speech on a CPU in 31 languages, Supertonic 3; for English preset voices under Apache 2.0, Kokoro; for small devices, Piper; for instant cloning on a GPU, Qwen3-TTS or VoxCPM 2.

Which free TTS models can be used commercially?

Of the models above: Supertonic 3 (OpenRAIL-M, with a list of prohibited uses), Kokoro, Qwen3-TTS and VoxCPM 2 (Apache 2.0) and Chatterbox (MIT). Piper's engine is GPL-3.0 and each voice has its own terms. The official OmniVoice, F5-TTS and XTTS-v2 weights are non-commercial, and Fish Audio S2 Pro needs a separate license.

Is there a free TTS model that runs without a GPU?

Yes. Supertonic 3, Kokoro, Piper and Chatterbox Nano run on an ordinary CPU. The larger cloning models also run on a CPU through community ports, but slower than real time.

Can free models clone a voice?

Qwen3-TTS, OmniVoice, VoxCPM 2, Chatterbox, F5-TTS and XTTS-v2 clone from a short clip; Kokoro and Piper do not. With Supertonic 3 a voice is trained from a recording into a file, for free on your own NVIDIA GPU or by an author to order.

GUIDE

The best free text-to-speech models for local use in 2026

Updated October 23, 2026

Free TTS is now good enough for real work, and many open models run on your own hardware. The word free still hides real differences: some licenses forbid commercial use, some models need a large GPU, and only a few run comfortably on an ordinary CPU.

This overview covers the open models we compare on this site. SuperClonica sells voices for one of them, Supertonic 3, so keep that in mind: we stick to facts from each model's own documentation and say plainly where another model is the better pick.

Listen to Supertonic 3 voices Install SuperClonica

How to choose a free TTS model

Four questions shorten the list quickly.

Language: does the model speak yours? Chinese, for example, rules out Supertonic 3.
License: free to download is not always free for commercial use. Check the terms of both the code and the weights.
Hardware: an ordinary CPU, or a GPU with enough memory?
Voices: are preset voices enough, or do you need a particular voice from a recording?

Small free TTS models that run on a CPU

Supertonic 3 from Supertone has about 99 million parameters, runs through ONNX Runtime and, according to its authors, needs no GPU. It speaks 31 languages, without Chinese, under OpenRAIL-M: royalty-free, with a list of prohibited uses. It ships with preset voices; more voices, free and paid, are in the SuperClonica catalog, and the SuperClonica engine runs them on average 61% faster on a CPU.

Kokoro-82M has 82 million parameters and an Apache 2.0 license, with 54 preset voices in eight languages. It is strongest in English: its own voice list rates the Mandarin and Japanese voices lower. The official release cannot clone voices. It runs in Python and in the browser through kokoro-js.

Piper is a fast local engine built for small devices like the Raspberry Pi and known as the voice of Home Assistant. Each voice is an ONNX model for one language; the list covers 39 languages, Chinese among them but not Japanese or Korean. The current engine is GPL-3.0, and every voice has its own license.

Chatterbox Nano from Resemble AI is an English model with 110 million parameters under MIT. Resemble AI reports 3× faster than real time on 8 CPU cores; it clones a voice from a clip and marks every file with an inaudible PerTh watermark.

Larger models for voice cloning

These clone a voice from a few seconds of audio. By our estimates they run slower than real time on a desktop CPU, so they are at their best on a GPU.

Qwen3-TTS: Apache 2.0, 0.6B and 1.7B sizes, 10 languages including Chinese, 3-second cloning and voice design from a description. About 1–2× slower than real time on a CPU (0.6B).
VoxCPM 2: Apache 2.0, 2B parameters, 30 languages plus Chinese dialects, 48 kHz audio, around 8 GB of VRAM. About 2–4× slower than real time on a CPU.
Chatterbox Multilingual V3: MIT, 500 million parameters, 23 languages, cloning from a clip; its examples run on a GPU.
Fish Audio S2 Pro: the top quality mark in our table and 80+ languages, but 4–10× slower than real time on a CPU, and commercial use needs a separate license from Fish Audio.

Free, but not for commercial use

Several popular models publish open code with non-commercial weights. They are fine for experiments and personal projects; a paid product needs a different model.

OmniVoice: 646 languages and cloning from 3–10 seconds of audio. The code is Apache 2.0, the pre-trained model CC-BY-NC.
F5-TTS: zero-shot cloning in Chinese and English. MIT code, CC-BY-NC weights, and the slowest CPU speed in our table.
Coqui XTTS-v2: 17 languages and cloning from a short clip. The Coqui Public Model License covers the generated audio as well, and the original repository is no longer maintained.

Which free text-to-speech model to pick

Each comparison page linked below has the sources and our speed estimates for its model.

Narration, games and apps on ordinary computers, in one of 31 languages: Supertonic 3, with a catalog voice or one trained from your recording.
English preset voices under the most permissive license: Kokoro.
Raspberry Pi and Home Assistant: Piper.
Instant cloning on a GPU with commercial use allowed: Qwen3-TTS, VoxCPM 2 or Chatterbox.
Chinese speech: Qwen3-TTS or VoxCPM 2 on a GPU; Piper and Kokoro also have Chinese voices.
Research and personal experiments: any of them, including OmniVoice, F5-TTS and XTTS-v2.