Kyutai releases Pocket TTS Small models with 1.9M to 18.2M parameters for CPU voice cloning

·2026.10.21 01:58

Key point

The new English models support voice cloning and are designed to run on CPU, allowing GPUs to remain dedicated to LLM inference.

Kyutai released three Pocket TTS Small models on October 20 with 1.9M, 6.9M, and 18.2M parameters.
The models are English-only, support voice cloning, and are optimized for CPU execution.
Running speech generation on CPU preserves GPU VRAM for LLMs that barely fit in memory.
Parameter counts include the audio decoder; the reference voice encoder is separate and only needs to be encoded once.
A demo is available to switch between model sizes, with code and setup instructions provided.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.