Underdog’s Woof speech model fits in 32MB and runs in a browser

Woof’s model card fills in the details. It speaks English in two voices, Juno and Mira, and reads prices, dates and phone numbers the way people say them. The two-voice download is 40.6MB and runs on a laptop processor using only numpy, a basic maths library, at 13.8 times real time with the Juno voice on one thread of an Apple M5 Max.

Underdog scored the model itself on 500 hard English sentences. By its count, Woof’s Mira voice gets 1.42% of words wrong, against 1.89% for Fish Audio Drama 3, a 41.5GB download, and 6.63% for ElevenLabs Flash v2.5. These are Underdog’s own figures.

The licence draws a clear line. Woof is free for personal, research, teaching and non-profit use if you credit Underdog, and always free inside Underdog’s own apps. An ad, a monetised video or a company’s work needs a commercial licence, with no public price.

Founder Sigil Wen told TechCrunch he wants open American-made models that keep data on people’s own devices.

On our AI voiceover page, ElevenLabs’ $6 Starter plan is the first to carry a commercial licence, and its Free plan covers personal use with attribution, our 8 November check.

For an advertiser, the cloud voice has a price you can read today: $6 a month on ElevenLabs Starter. Woof’s local voice needs an email to Underdog before it can go into an ad.