Pocket TTS Shrinks to 1.9M Parameters: Voice Clone on One CPU Core
Kyutai Labs just dropped three new distilled versions of Pocket TTS, and the smallest is almost too good to believe. A 1.9M-parameter model that runs 103x faster than real-time on one CPU core, alongside a 6.9M model and an 18.2M model with the quality of the original. All three still support voice cloning.
What Changed
Back in January, Pocket TTS arrived as a 100-million-parameter open-source TTS model that runs on CPUs in real-time and supports voice cloning. That was tiny. This is absurd. Kyutai distilled a ~300M teacher, then shrank the student in steps: cut the model, retrain it, cut again. The result is something that fits on a phone, in a browser tab, on a Raspberry Pi—and still speaks naturally.
What People Are Building With It
Developers are already using Pocket TTS for offline voice assistants, game voiceovers, and accessibility tools. Real-time applications like live voice assistants, interactive gaming, and dynamic audiobook narration are now possible because the model can handle infinitely long text without memory explosion. The Sonorus mod for Hogwarts Legacy demonstrates this: players talk to any named character in their original voice. One community builder made a real-time audio synthesis tool for games and apps with voice cloning based on user presets.
With the new tiny models, the possibilities shift from "runs on your laptop" to "runs embedded in your app." A single 64 MB download (the smallest with quantization) means voice can go places it never could before.
What's Actually Good (and What Isn't)
The speed is real. The model runs on CPU with strong single-thread performance (Apple Silicon), and uses only 2 CPU cores. On a MacBook Air M4, it achieves 6x real-time throughput with ~200ms latency to get the first audio chunk. The 1.9M model at 103x real-time means you can generate an hour of audio in about 35 seconds on a single core.
Voice cloning works, but it needs a reference. The model can clone a voice from a short audio sample of around five seconds, capturing tonal qualities, accent, and acoustic characteristics, and cloned speech maintains a high level of similarity to the reference voice. The trade-off is predictable: smaller models trade quality for speed. The 18.2M variant preserves most of the original 100M quality. The 6.9M is noticeably thinner. The 1.9M model works, but you'll notice.
How to Try It
All three models are on GitHub and Hugging Face under MIT license, free to use. Installation is a single pip command, then you run the CLI or Python API. Support includes English, French, German, Portuguese, Italian, Spanish, and Dutch. Pass your own audio file as a voice reference to clone. The community has already built browser screen readers and other projects on top of it.
If you're building voice into a product and don't want to depend on cloud APIs or proprietary models, this is the moment to try locally. The 18.2M model is the practical sweet spot—most people won't hear the difference, and it runs everywhere.