Pocket TTS
Kyutai's 100M voice-cloning speech model that runs faster than real time on a laptop CPU. Six languages, CC BY 4.0 weights behind a gate.
In short. Pocket TTS reads text aloud and can copy a voice from about five seconds of audio. It is small enough to run faster than real time on an ordinary laptop processor, with no graphics card. The weights are free to use with credit to Kyutai, but you must accept a notice that forbids cloning anyone's voice without consent.
Summary
Pocket TTS is a text-to-speech model from Kyutai, a non-profit research lab in Paris, announced on 13 January 2026. It has about 100 million parameters, which Kyutai splits into a 90-million-parameter generator and a 10-million-parameter audio decoder, plus an 18-million-parameter encoder that runs once for each voice sample. Instead of predicting discrete audio tokens, it predicts continuous audio features one 80-millisecond frame at a time and decodes them with a decoder derived from the Mimi codec. The method comes from Kyutai's Continuous Audio Language Models paper. According to the model card, it clones a voice from about five seconds of audio, starts speaking after roughly 200 ms, and runs about six times faster than real time on a MacBook Air M4 using two CPU cores. The card lists English, French, German, Portuguese, Italian and Spanish, and the GitHub README also lists Dutch.
The launch post tested it on LibriSpeech test-clean against F5-TTS, Kyutai's own 1.6B TTS, Chatterbox Turbo and Kokoro. Pocket TTS had the lowest word error rate, 1.84%, tied with the 1.6B model. In human pairwise ratings it beat F5-TTS and the 1.6B model on audio quality but trailed Chatterbox Turbo, and its speaker similarity was the lowest of the four voice-cloning models, with an Elo of 1,898 against 1,946 to 2,037. It was the only voice-cloning model in the comparison that ran faster than real time on the test laptops' CPUs. The first English model was trained only on about 88,000 hours of public English datasets. These are the developer's own results.
Kyutai has kept extending it. The repository now has per-language models, with larger 24-layer variants for the non-English languages, and the training code was released in August 2026. In October 2026 Kyutai published much smaller English models of 18.2, 6.9 and 1.9 million parameters, decoder included, made by shrinking a 300-million-parameter teacher. On its LibriSpeech test, the 18.2M model scored a word error rate of 0.85% against 0.92% for the released 100M model, and ran at 9.9 times real time on one CPU thread against 2.8 times. The 1.9M model ran at 25 times real time with a 1.72% error rate. These are also Kyutai's own measurements.
The license needs care. The code is MIT. The weights on the Hugging Face repository are listed as CC BY 4.0, although the launch post says Pocket TTS was released under the MIT license, so check the repository you download from. That repository is gated: you accept a notice that bans voice impersonation or cloning without explicit and lawful consent, misinformation and deception, and unlawful or harmful content. The preset voices come from several datasets, each with its own license, listed in Kyutai's kyutai/tts-voices repository. A sibling repository, kyutai/pocket-tts-without-voice-cloning, publishes weights without the cloning feature and is not gated.
Good for
- Cloning a voice from a short sample on a laptop or a small server with no GPU
- Offline, private read-aloud, assistants and games, since audio streams after about 200 ms
- Research and fine-tuning, because Kyutai released the training code and trained the first model only on public datasets
- Very small deployments, using the October 2026 English models of 18 million parameters or fewer
Watch out for
- The weights are listed as CC BY 4.0 and sit behind a prohibited-use gate, while the launch post says MIT, so confirm the terms of the repository you use
- Each preset voice has its own license, so check it before shipping a voice in a product
- In Kyutai's tests its speaker similarity trailed Chatterbox Turbo and Kyutai's own larger model
- Cloning reproduces the quality of the sample, including noise, so clean the audio first
- Pauses cannot be inserted with silence in the text, and GPU use is not officially supported
- The speed and quality figures are the developer's own
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Pocket TTS. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Pocket TTS 18M (English only) | 18M | 0.11 GB | 0.127 GB | Laptop or phone |
| Pocket TTS 100M (six languages) | 100M | 0.22 GB | 0.253 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
A Creative Commons license that lets anyone copy, share and adapt the work for any purpose, including commercial use, as long as they give credit. It was written for creative and data works rather than software, but some speech and data releases use it for model weights.
- Attribution or model name must be shown
See also
- Kokoro vs Pocket TTS: which tiny open TTS model should you use?
- Open-weight AI models made in Europe
- Best open-source speech models for transcription and text-to-speech
- Best open-source text-to-speech models in 2026
- Open speech models: transcription is open, text-to-speech is mixed
- Tiny text-to-speech models that run on a CPU: four open options compared
- ZONOS3-Mini: open TTS in 43 languages that runs in a browser
Frequently asked questions
- What is Pocket TTS?
- Pocket TTS reads text aloud and can copy a voice from about five seconds of audio. It is small enough to run faster than real time on an ordinary laptop processor, with no graphics card. The weights are free to use with credit to Kyutai, but you must accept a notice that forbids cloning anyone's voice without consent.
- Can I use Pocket TTS commercially?
- Yes. Pocket TTS is released under CC BY 4.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Pocket TTS open source?
- Pocket TTS is open-weight, not open source in the strict sense. You can download the model files, but the CC BY 4.0 adds conditions that the Open Source Initiative definition does not allow.
- Can I run Pocket TTS on my own computer?
- Yes. The smallest variant, Pocket TTS 18M (English only), needs about 0.127 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Pocket TTS released?
- Pocket TTS was released by Kyutai on January 13, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Kyutai: Pocket TTS, a high-quality TTS with voice cloning that runs on CPU (13 Jan 2026) kyutai.org
- Kyutai: The tiny TTS was actually 10x too big (20 Oct 2026) kyutai.org
- Pocket TTS repository and README (GitHub) github.com
- Pocket TTS model card and prohibited-use notice (Hugging Face) huggingface.co
- Continuous Audio Language Models (arXiv 2509.06926) arxiv.org
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Phonon 31M Gradium | 31M | – | Apache 2.0 (Permissive) | 0.0747 GB Laptop or phone | |
| Kokoro hexgrad | 82M | – | Apache 2.0 (Permissive) | 0.38 GB Laptop or phone | |
| Supertonic 3 Supertone | 99M | – | OpenRAIL-M (Conditional) | 0.46 GB Laptop or phone | |
| MOSS-TTS-Nano OpenMOSS | 100M | – | Apache 2.0 (Permissive) | 0.276 GB Laptop or phone |