Kyutai open-sources Pocket TTS stack, trains for under $200
syhw · x · 2026-08-26
Kyutai Labs has open-sourced the entire Pocket TTS training stack, including data pipelines, recipes, and evals. Designed for CPUs, it aims to run TTS on any device via a simple pip install.
Training is fast: 15k steps yields words, 50k steps reads anything (WER under 1%), and 200k steps removes synthetic sounds. It takes a week on a consumer GPU or 10-20 hours on 8 H100s. Renting hardware costs less than $200 per run.
The team encourages users to train in their own language (a few hundred hours suffices), add features like emotion or singing, or optimize for speed and size.
More from coding & agent
- Train a private SLM on iMessages locally with MLX — romainhuet · 2026-08-26
- Dev builds a playable Game Boy Advance shooter 'Sunshard' with Codex — romainhuet · 2026-08-26
- Claude Reflect System: Enables permanent learning from corrections — tom_doerr · 2026-08-26
- Bug Hunt Bench: Benchmarking frontier coding models on real planted bugs — PawelHuryn · 2026-08-26
- Built a Lego Instructions App using Grok's Build Mode — Baconbrix · 2026-08-26
- JetBrains Releases Junie Local, Bringing Its Coding Agent Fully On-Device to Macs — homothebrave · 2026-08-26