Underdog demos 31.7MB Woof voice model in a browser game

Underdog puts its compact voice model into an interactive demo

Underdog has released a retro-themed browser game to demonstrate Woof, its new frontier text-to-speech model. The model is 31.7MB in size, has 31 million parameters, and can speak text entered through a virtual typewriter inside the game.

The demo presents a room with a CRT monitor, desk lamp, typewriter and music tuner. Users can type with their keyboard and hear the generated speech, while rotating the tuner produces different kinds of music. Underdog says all voices in the game are generated by Woof.

The company positions the project against cloud-dependent AI services: the game’s introduction refers to “intrusive” cloud providers. Underdog says Woof works in a browser and can run on most devices without an internet connection, making the demo an example of local speech generation rather than a remote API experience.

Open model aimed at device-based AI

Woof is available for anyone to download. Underdog does not present the small model as a replacement for much larger systems, but the release points to uses such as local assistants and personalised podcast creation where a compact model may be sufficient.

The launch follows Bark, Underdog’s 167MB speech-to-text model for edge-device transcription and dictation. The company has also released a 27-billion-parameter model that it says can run on laptops, as well as a waitlisted locally hosted application positioned as an alternative to Muse and Instinct.

Founder Sigil Wen said he began building the application after wanting a personal AI note-taking system that would run on users’ devices and leave users in control of their data. That direction aligns with Underdog local AI assistant launch and its focus on a locally run AI assistant, while Woof extends the company’s product work into speech generation.

What the demonstration means for teams

Woof illustrates a deployment choice rather than a claim that model size no longer matters. Larger systems may remain necessary for tasks that require broader capabilities, but a 31.7MB voice model changes the practical threshold for features that can run close to the user.

For businesses evaluating voice interfaces, the practical implication is to identify narrow speech tasks that can operate on existing devices, then test local models against requirements for quality, offline operation and data control.