Underdog demos Woof, a 32 MB text-to-speech model that runs in the browser

Underdog, a startup building AI models meant to run on users' own devices, has released Woof, a small text-to-speech model, and showcased it in a game that runs entirely in the browser. TechCrunch reported the release on 6 November, after the company shared the demo a day earlier. The version used in the browser is 31.7 MB, which explains the "32 MB" figure in coverage.

What Woof does

According to the model card published on Hugging Face under the Conway Research account, Woof is an English-only speech synthesis model with two voices, Juno and Mira, producing 24 kHz mono audio. A built-in rule-based normaliser turns numbers, dates, times, amounts of money and codes into spoken words. The full two-voice release weighs 40.6 MB, and the card lists about 38.2 million parameters, split between a speech model and an audio decoder adapted from Kyutai's Mimi codec. TechCrunch cited a figure of 31 million parameters, a discrepancy the company has not explained publicly.

The card says the model runs on an ordinary CPU using numpy, without a GPU, server or API key. Underdog reports that the Juno voice generates audio 13.8 times faster than real time on a single Apple M5 Max CPU thread, and gives word error rates of 1.42% (Mira) and 1.53% (Juno) on one English test set. These are the company's own measurements; FourWeekMBA notes they have not been independently reproduced.

The browser demo

The showcase is a retro-themed game set in a room with a typewriter, a CRT monitor and a desk lamp. Visitors type on the typewriter and the game reads the text aloud, and TechCrunch reports that turning a tuner produces different styles of music. The model card itself describes Woof only as a text-to-speech model, so it is unclear which component generates the music.

Licence and context

Woof is downloadable, but not open source in the usual sense. The Underdog Non-Commercial License 1.0 allows personal, research, teaching and non-profit use with credit, requires a separate licence for commercial use, and forbids using Woof or its output to train or distil other speech models.

Size is the main selling point. FourWeekMBA calculates that Woof's file is roughly eight times smaller than Kokoro-82M's, while Underdog's card acknowledges that Kitten Nano is smaller but less accurate. Founder Sigil Wen told TechCrunch he expects most of today's capable models to run on users' existing devices within months, a prediction rather than a demonstrated result. The company previously released Bark, a 167 MB speech-to-text model.