Kyutai · 2026 · on-device clone

Kyutai Pocket TTS

A ~100-million-parameter Kyutai TTS, small enough to clone a voice and run in real time on CPU. Related to their delayed-streams work, not the Moshi dialogue model. The clips here are the shared bench scripts, generated locally with the official alba voice.

Known for

Open voice cloning that actually fits in a laptop CPU budget.

Shared bench script

alba · Generated locally with Kyutai Pocket TTS, official alba voice prompt.

Listen closely. Every voice on this bench is a machine, reading the same line.

Shared bench script

alba · Generated locally with Kyutai Pocket TTS, official alba voice prompt.

The rain had not stopped for three days. On the fourth morning, the river finally broke its banks, and the town of Millfield learned what water can do when it is given time.

Shared bench script

alba · Generated locally with Kyutai Pocket TTS, official alba voice prompt.

I told you we should have left at dawn. And miss the last train? Don't be ridiculous. Sit down. We'll figure this out.

Shared bench script

alba · Generated locally with Kyutai Pocket TTS, official alba voice prompt.

At 8:15 this morning, officials confirmed that Route 47 will remain closed through Friday while crews repair the collapsed overpass near Cedar Junction.

Shared bench script

alba · Generated locally with Kyutai Pocket TTS, official alba voice prompt.

Please transfer $1,247.50 to account 883-21, then call Dr. Nguyen at 555-0142 before 3:30 p.m. on March 9th.

Approach
Streaming
Year / era
2026
License class
open
Cloning
Yes
Streaming
Yes
Emotion controls
No
Languages (tags)
en · fr · multilingual
Official
Source
Kokoro 82M · af_heart