Pocket TTS Shrinks to 1.9M Parameters
Published 23 October 2026
Kyutai presents smaller Pocket TTS models, reducing the system from roughly 100M to as little as 1.9M parameters while retaining real-time CPU voice cloning. The open-source techniques include gradual backbone pruning, head and decoder distillation, and optional reward fine-tuning.
Read the original on kyutai.org →