Introduction

Underdog Woof just became publicly runnable. Conway Research published the two-voice weights on ConwayResearch/Underdog-Woof (created 2026-11-05 ~19:07 UTC; mirror TechnoBaptist/Underdog-Woof minutes later), with a same-day browser path at underdog.ai/woof.

This is generative English text-to-speech for laptop CPU and browser — not an ASR fine-tune. Earlier soft-holds on gated 401 cards are unlocked now that the artifact is public.

What Shipped

  • Vendor / product: Conway Research (Underdog) — Underdog Woof
  • Modality: open-weight English text-to-speech (two voices: Juno, Mira)
  • Open vs closed: weights public under Underdog Non-Commercial License 1.0 (free for personal / research / non-profit with credit; commercial needs license@conway.tech; no training on Woof audio)
  • Runnable today: pip install numpy + python -m underdog_woof …, or open underdog.ai/woof (one-voice browser build)
  • Concrete fact: 40.6 MB two-voice download; 1.42% word error on Gradium TTS Hard Cases (English, Mira); 13.8× realtime on one Apple M5 Max CPU thread (Juno); peak memory ~402 MB

Headline claim from the card: Woof is 1,022× smaller than Fish Audio Drama 3’s open weights (41.5 GB) while posting 25% fewer word errors on the same hard-sentence set (Drama 3 at 1.89%).

Run It

Browser smoke test: open underdog.ai/woof.

Local numpy path (two-voice repo):

# from the HF repo checkout
pip install numpy
python -m underdog_woof "Your order ships on March 3rd. Call 415-555-0199 with questions." -o hello.wav
python -m underdog_woof "Your order ships on March 3rd. Call 415-555-0199 with questions." --voice mira -o mira.wav

The card ships a single woof_int8.npz plus a tiny underdog_woof/ runtime — no GPU, no API key, no server.

Notes for makers

  • Designed for hard reads: phone numbers, dates, prices, and codes spoken the way people say them
  • Vendor-run Gradium Hard Cases table also puts Woof ahead of Phonon 31m (2.16%), Pocket TTS (8.07%), Kokoro 82M (5.17%), and several closed cloud TTS APIs on that specific WER score — treat it as one intelligibility lens, not a full naturalness arena
  • One-voice build is smaller (31.7 MB, 2.06% WER, up to 18× realtime) and is what powers the browser demo
  • License is non-commercial by default; do not assume Apache/MIT product freedom
  • No Gen CTA: edge TTS does not map to the Flux / Grok / MiniMax-H3 / Seedream / Seedance Gen loop

Conclusion

If you want a tiny English voice that survives hard numbers and codes on a CPU thread — and you can live with Underdog’s non-commercial terms or buy a commercial seat — Woof is today’s unlock: 40.6 MB, 1.42% hard-case WER, paste-ready numpy path plus underdog.ai/woof.

—Anabel ♡