Kyutai releases smaller Pocket TTS versions with up to 1.9M parameters

What happened and what to do

Kyutai Labs states, in the announcement, that it has released smaller versions of Pocket TTS, derived from the original model of about 100M parameters. The new variants include models of 18.2M, 6.9M and 1.9M parameters. According to the author, the 1.9M version runs 103 times faster than real time on a single CPU core and keeps voice cloning. The text describes the project as a small, self-hostable speech synthesis model aimed at on-device voice applications.

For a company, a model like this makes it possible to evaluate speech synthesis without depending on a GPU or an external API. In practice, the first step is to measure latency, memory use and audio quality on the actual equipment where the voice will run, such as terminals, kiosks or edge servers. A policy for voice cloning is also needed, with consent records, access control and an audit of who generated each voice. The solution can then be packaged on the client's infrastructure, with performance monitoring and a process for updating model versions.

How the consultancy can help

Feasibility diagnosis for local voice: latency and quality measurement on the company's hardware, design of a self-hosted TTS stack with voice cloning and consent controls, and deployment with monitoring on the client's infrastructure.

Next step

If your company wants to evaluate its own speech synthesis on CPU, send a short description of the case, such as the channel of use, expected volume and available hardware, and receive a scoped proposal.