Kyutai has released smaller versions of its Pocket TTS speech model, down to 1.9 million parameters while keeping voice cloning. The original model had about 100 million parameters and already ran in real time on a CPU.
The family now includes three variants. The largest, at 18.2 million parameters, matches the quality of the original. An intermediate 6.9 million model aims at a balance between size and output. The smallest, at 1.9 million, is the most striking: according to Kyutai, it runs 103 times faster than real time on a single CPU core.
Cloning works from a sample of about 5 seconds of audio. The model reproduces timbre, emotion, accent, cadence and even the acoustic conditions of the recording, such as reverb and microphone. The code is on GitHub, and the project can be tried in the browser through services such as OfflineTTS, which let users record or upload a clip to build a local copy of a voice.
For people following local TTS, the key number is the trade-off between size and performance. At 1.9 million parameters, speech synthesis becomes practical on devices with minimal resources, without relying on a cloud service. A similar path had already emerged with Sopro, which focused on cleaner cloned voices on a lightweight model, while Gradium chose open models under 50 million parameters built on Pocket TTS and Kyutai's Mimi codec.
The same ease of cloning, however, reopens the question of consent. The issue is already central to the debate around local voice cloning with Qwen3-TTS, where a few seconds of audio are enough to reproduce a person.
No comments yet. Be the first!