31.7MB open-weight TTS model Woof beats 1000x-larger Fish Audio Drama 3 on WER
alexcovo_eth · x · 2026-11-07
Underdog released Woof, an open-weight TTS model in two sizes: 31.7 MB (one voice) and 40.6 MB (two voices), small enough to run fully in-browser offline and up to 18x real-time on a single MacBook CPU thread.
On a 500-sentence test packed with phone numbers, dates, prices and codes, the two-voice version scores 1.42% WER — 25% fewer errors than Fish Audio Drama 3 (1.89%, 41.5 GB download, 35.9B parameters) at roughly 1/1000th the size. The one-voice model scores 2.06%, about on par with Drama 3. The team's benchmark chart also compares Kokoro, NeuTTS Air, Qwen3-TTS, Sesame CSM and other open voice models.
Weights are open under CC BY-NC 4.0 (non-commercial), with the two-voice model on Hugging Face. Numbers are self-reported.
Related event: Tiny 31.7MB Open-Source TTS Model Woof Beats Fish Audio Drama 3(2 posts)→
More from Infra
- Dev mocks anti-data-center crowd: where do you think your tweets come from? — cgarciae88 · 2026-11-07
- Nvidia cleared to sell H200 chips to an empty office in China, NYT finds — The Verge AI · 2026-11-07
- Microsoft engineer built a tool that reveals what's really using your GPU — BrooklynBrawl · 2026-11-07
- Starship's first orbit beats 20 Falcon 9 launches: inside SpaceX's five-business strategy — hey_abusiddik · 2026-11-07
- Google's Willow chip: a 5-minute calculation, parallel-universe hype, and the real science — hey_abusiddik · 2026-11-07
- The Agent Cloud is actually two clouds: one where models think, one where agents work — AccBalanced · 2026-11-07