A 4.5M-param, 21MB TTS model trained for $18 runs in your browser
victormustar · x · 2026-11-16
Caleb Fahlgren trained a tiny text-to-speech model on top of tiny-voca (by @n4ze3m): just 4.5M parameters and 21MB, running in the browser (even on CPU). It learned by listening to 25 hours of Maya-1 — a model 727x its size — so it can sigh, exhale and gasp. Total training cost: $18.
More from Multimodal
- Same Prompt, Two Groks: Grok Chat vs Grok Imagine Video Generation Compared — minchoi · 2026-11-17
- Seedance 2.5 Turns Grey 3D Blockouts Into Finished Anime Shots—No MoCap Needed — minchoi · 2026-11-17
- A World Series Home Run Rebuilt as a Gaussian Splat, Watchable From the Catcher's Eyes — minchoi · 2026-11-17
- Full Room Gaussian Splat Relit in Blender 5.3 With Cycles Using Custom Delighting — minchoi · 2026-11-17
- $250 robot arm hands you tools on command with zero training, using Meta's Muse Spark + SAM 3.1 — minchoi · 2026-11-17
- A $250 robot hands you tools, fly-around World Series home runs: 10 wild AI visual demos this week — minchoi · 2026-11-17