Underdog Woof: 32MB text-to-speech, non-commercial license

Underdog, the on-device AI startup founded by Sigil Wen, showed off Woof this week: a text-to-speech model small enough to run in a browser tab. The browser version is 31.7MB. The two-voice download on Hugging Face is 40.6MB and runs on a plain CPU, with no GPU and no API key. For anyone who has paid per character for a voice API, that's a real shift. There's one catch, and it's in the license file, not the demo. Woof's weights are free to download. They are not free to use in a product you sell.

What actually happened

TechCrunch reported on November 6 that Underdog shipped Woof inside a browser game, where text typed on a virtual typewriter is spoken back. The company says the model can run on most devices without an internet connection. The details below come from the Woof model card, published under Underdog's Conway Research organization:

• Size. 38.24M parameters for the two-voice model (Juno and Mira), 40.6MB on disk. The single-voice browser build is 31.7MB.
• Output. English only, 24 kHz mono audio.
• Runtime. CPU with numpy, from a single weights file. No hosted API.
• Benchmarks. On the "TTS Hard Cases" English set, the card lists word error rates of 1.42% (Mira) and 1.53% (Juno). On one Apple M5 Max CPU thread, Juno runs 13.8x faster than real time. These are Underdog's own numbers.
• License. The Underdog Non-Commercial License 1.0. Personal, research, teaching and non-profit use is allowed with credit. Commercial use needs a paid license from the company. You also can't use Woof to train or distill other speech models.

Underdog shipped a speech-to-text model, Bark 3.0, a few weeks earlier: 163MB, English, under CC BY-NC 4.0. Same pattern, also non-commercial. We covered Underdog's on-device assistant at its October launch.

Why Woof matters for your business

A 40MB voice that runs on the CPU means read-aloud order confirmations, kiosk prompts and accessibility audio with no per-character bill and no customer text leaving the device. That's the direction small models are heading, and it's good news for operators.

But "open to download" and "open source" are different promises. If your customer-facing app speaks with Woof and you haven't bought a license, you're out of terms. Many weekend prototypes go to production this way.

So prototype with it, and price it. Email for the commercial terms before you build a feature on it. Then compare against permissively licensed small voice models and your current API bill. Keep the voice layer behind one interface in your code so you can swap it when the terms or a better model change.

Sources: TechCrunch, Underdog Woof model card, Underdog Bark 3.0 model card.