Woof

Woof is Underdog's (weights under ConwayResearch) tiny English text-to-speech model: the one-voice build is about 31.7MB / 30.7M parameters and runs locally in-browser via WebGPU (claimed offline-capable); a two-voice package is ~40.6MB with fixed voices Juno and Mira. Demo: https://underdog.ai/woof ; TechCrunch: https://techcrunch.com/2026/11/06/underdog-shows-off-a-32mb-speech-model-in-new-demo/ .

Ship notes (demo 2026-11-05 / coverage 11-06):

• Packaged as a retro-office browser game — type on a virtual typewriter and hear local speech;
• 24 kHz mono output with a built-in rule-based text normalizer;
• Open weights under non-commercial terms (see official repo license);
• Sits next to Underdog's Bark STT in a local-first product story.

Why it matters

Cloud TTS dominates; Woof packages a downloadable, browser-runnable speech model into a playable demo and pushes the size narrative to ~32MB — useful for private/offline assistant prototypes.

Caveats

• Fixed voices; this release is not a voice-cloning product;
• Vendor real-time speed claims need independent verification; WebGPU cold-start can add latency;
• Distinguish from earlier Underdog namesakes — use ConwayResearch/Underdog-Woof and underdog.ai/woof.

1. Try the browser demo at https://underdog.ai/woof ;
2. Download weights from ConwayResearch/Underdog-Woof on Hugging Face;
3. Follow the repo for numpy/CPU or WebGPU integration;
4. Check the non-commercial license before production use.