4.5M-parameter TTS model runs 70x real-time on CPU with just $10 training cost

mrkn1 · hn · 2026-11-18

tiny-voca, an open-source TTS model on Hugging Face, packs only 4.5M parameters yet runs 70x real-time on CPU and cost about $10 to train — making on-device and low-cost speech generation practical.

Original post →

More from Infra

Infra channel →