Kyutai Pocket TTS
Kyutai
Kyutai Pocket TTS is a 100-million-parameter text-to-speech model that the vendor says was released in January 2026 and runs in real time on a CPU using only two cores, reaching roughly six times real time on a MacBook Air M4. It streams audio with about 200 ms to the first chunk, clones a voice from any audio sample locally, ships a catalogue of preset voices, and handles arbitrarily long text inputs. It speaks English, French, German, Spanish, Portuguese and Italian, and can run client-side in the browser.
Strengths
- Only 100 million parameters, running faster than real time on a CPU (about six times real time on a MacBook Air M4) using just two cores
- Audio streaming with roughly 200 ms to the first audio chunk
- Voice cloning from any audio sample when run locally, plus a published catalogue of preset voices with per-voice licences
- Six languages: English, French, German, Spanish, Portuguese and Italian
- Handles infinitely long text inputs, and can run client-side in the browser
- Python API, CLI (generate, serve, export-voice) and a local HTTP server, installable with a single pip or uvx command
- Training code released in August 2026 so the model can be retrained on new data
Caveats
- Kyutai states only the month, 'Released in January 2026', so no exact release date is recorded here rather than inventing a day.
- The weights are CC BY-4.0 and gated on Hugging Face behind a contact-information form and a prohibited-use agreement, even though the code is MIT.
- Six languages only, and the repository says the non-English languages have 'biggers 24 layers variants that are higher quality but slower', implying the default non-English models trade quality for speed.
- No accuracy, intelligibility or listener-preference benchmarks are published on the vendor's TTS page or in the repository; only speed and latency figures are given.
- Cloning quality is bounded by the input sample: the repository warns that 'the audio quality of the sample is also reproduced' and recommends cleaning samples first, and the Hugging Face gate explicitly prohibits voice cloning without lawful consent.
- Vendor
- Kyutai
- License
- CC-BY-4.0
- Modalities
- textaudio
Capabilities
Resources
Pricing
Kyutai publishes no hosted API or pricing. The weights are distributed on Hugging Face under CC BY-4.0 behind an access-acceptance form, and the pocket-tts code is on GitHub under the MIT licence, for local CPU inference. A free browser demo is hosted on kyutai.org.