Tiny text-to-speech models that run on a CPU: four open options compared

Kokoro, Pocket TTS, Phonon 31M and Supertonic 3 are open text-to-speech models of 100M parameters or fewer that run on a laptop CPU. Compare sizes, languages, voice cloning and licenses.

  • Analysis
  • 4 min read
  • Published

Four open text-to-speech models of 100 million parameters or fewer now run on an ordinary laptop CPU: Kokoro, Pocket TTS, Phonon 31M and Supertonic 3. Kyutai’s Pocket TTS arrived in January 2026, Supertone’s Supertonic 3 in May and Gradium’s Phonon in October, while Kokoro has been out since January 2025. They look alike on a spec sheet and differ in ways that decide which one you can ship.

The four at a glance

Model Size Languages Voice cloning License of the weights
Kokoro (hexgrad) 82M, 327 MB 8, with 54 voices No Apache 2.0
Pocket TTS (Kyutai) 100M, 0.22 GB 6, plus Dutch in the README Yes, from about 5 seconds CC BY 4.0 behind a prohibited-use gate
Phonon 31M (Gradium) 31M, 64.5 MB English only No, API only Apache 2.0
Supertonic 3 (Supertone) 99M, about 400 MB 31, no Chinese No, 10 preset voices OpenRAIL-M

Speed and quality are self-reported

Every number below comes from the developer of that model, run on different hardware and test sets. Kyutai reports that Pocket TTS runs about six times faster than real time on a MacBook Air M4 using two CPU cores, and that it and Kokoro were the only models in its laptop comparison faster than real time on a CPU. Kyutai’s October update adds English models of 18.2M and 1.9M parameters, running at 9.9 and 25 times real time on one CPU thread. Gradium says Phonon 31M starts speaking in about 7 ms on a laptop CPU. Supertone says Supertonic 3 runs fast on a CPU even compared with larger models measured on an A100.

Quality claims do not line up either. In Gradium’s English test, Phonon 31M scores an Elo of 2,098 against 2,043 for Kokoro. In Kyutai’s tests Pocket TTS has the lowest word error rate, 1.84%, but trails Chatterbox Turbo on audio quality and has the lowest speaker similarity of the four voice-cloning models. Listen to your own text before you choose.

Licenses decide more than size

Only Kokoro and Phonon 31M use plain Apache 2.0 for their weights. Pocket TTS is listed as CC BY 4.0 on Hugging Face behind a notice that bans impersonation, deception and unlawful content, while Kyutai’s launch post says MIT, so check the repository you download from. Each preset voice has its own license, and a sibling repository without the cloning feature is not gated. Supertonic 3 uses OpenRAIL-M, an OpenRAIL license that allows commercial use but requires every copy and derivative to carry its list of banned uses. Phonon’s larger 40M models are CC BY-NC and not free for commercial use.

What you give up

Small models trade away control. Kokoro’s voices are fixed, most of them work best on text of roughly 100 to 200 tokens, and long text should be split into chunks. Pocket TTS cannot insert pauses with silence in the text, and GPU use is not officially supported. Phonon 31M splits long input into sentences, so intonation does not carry across the whole text, and it reads other languages with English pronunciation. Supertonic 3 has ten preset voices, and cloning was offered only through Supertone’s hosted services. If you need long-form narration or many custom voices, a larger model on a GPU is the better fit.

Which to pick

  • The smallest file, in English: Phonon 31M is 64.5 MB, has 37 voices and a runtime for Python, Rust, C and C++, Swift, Android and the browser.
  • Voice cloning on a CPU: Pocket TTS clones a voice in six languages, but read the gate and the voice licenses first.
  • The most widely used: Kokoro had about 12.5 million downloads in the 30 days to 11 November 2026, with the broadest set of ready-made voices. It cannot clone a voice, and its last model release was in February 2025.
  • Many languages, including Korean, Japanese and Arabic: Supertonic 3, but Supertone archived the project in September 2026, so expect no fixes.
  • A larger model that still runs locally: ZONOS3-Mini has 800M parameters but only about 100M active, clones voices in 43 languages and, Zyphra says, runs in real time on laptops, phones and in a browser. It is Apache 2.0 and was released on 9 November 2026, so it has little track record.
  • Two more CPU options outside this comparison: MOSS-TTS-Nano from OpenMOSS has about 100M parameters, clones voices, lists around 20 languages and writes 48 kHz stereo, and OpenMOSS says its ONNX version ran smoothly on one core of a MacBook Air M4. VieNeu-TTS v3 is a 131M model for Vietnamese and English with 23 voices and cloning, and its card puts first audio at about 300 ms on a CPU in fp32. Both are Apache 2.0, and neither project publishes a quality benchmark for its speech model.

The best open-source text-to-speech models list ranks the larger options too, and how to choose a text-to-speech model walks through the decision. Voice cloning explains the consent rules these licenses refer to.

Sources

Models in this story

Model Released Parameters Context License Smallest variant needs
Kokoro hexgrad 82M – Apache 2.0 (Permissive) 0.38 GB Laptop or phone
Pocket TTS Kyutai 18M – 100M – CC BY 4.0 (Permissive) 0.127 GB Laptop or phone
Phonon 31M Gradium 31M – Apache 2.0 (Permissive) 0.0747 GB Laptop or phone
Supertonic 3 Supertone 99M – OpenRAIL-M (Conditional) 0.46 GB Laptop or phone
ZONOS3-Mini Zyphra 800M 100M active – Apache 2.0 (Permissive) 10 GB Laptop or phone
MOSS-TTS-Nano OpenMOSS 100M – Apache 2.0 (Permissive) 0.276 GB Laptop or phone
VieNeu-TTS v3 Phạm Nguyễn Ngọc Bảo (VieNeu-TTS) 48M – 131M – Apache 2.0 (Permissive) 0.322 GB Laptop or phone