Underdog releases its 32 MB Woof speech model

Startup Underdog introduced Woof, an open speech synthesis model measuring 31.7 MB that can run in a browser and, according to the company, on most devices without an internet connection. To introduce it, the startup released a game: players type text on a typewriter, and Woof reads it aloud; turning a music tuner changes the sound of the music. The model has 31 million parameters. Underdog has also previously released Bark, a 167 MB speech recognition model, and a model with 27 billion parameters that runs on laptops.

Underdog releases 32 MB Woof speech model

A game to test Woof

The game takes place in a room filled with retro technology: there is a typewriter, a cathode-ray tube monitor and a desk lamp. Clicking its switch makes the lamp produce a sound. In the opening, players are warned about intrusive cloud providers in the city — apparently a dig at cloud models.

Players can type any text on the typewriter, and the game will read it aloud. The music tuner lets them turn a dial to change the types of music. According to a post by Underdog on X, Woof created all the voices in the game. To get an invitation, players have to defeat Darkcloud.

The model runs directly in the browser. Underdog says Woof works on most devices without internet access and invites anyone interested to download it.

A bet on models for personal devices

A small model is unlikely to replace larger ones, but it could work for local voice assistants and personalized podcasts. A week or months earlier, Underdog released Bark, a 167 MB speech recognition model for resource-constrained devices. It can be used for transcription and dictation.

Underdog founder Sigil Wen told TechCrunch that he started building an app because he wanted his own note-taking system. Rather than use ready-made tools such as Granola, he wanted to use AI models running on users’ devices, so they could control their own data.

Wen said he hoped that within a few months, most sufficiently powerful modern models would be able to run on devices people already own. He said Underdog wants to train open models made in the US that are capable enough to let people use AI while preserving their privacy. He connected these goals to his manifesto.

In addition to Woof and Bark, the company has released a model with 27 billion parameters that can run on laptops. Underdog also has an app with a waitlist; the company positions it as a local alternative to Muse and Instinct.

Dataist's take: Woof shows how a compact speech synthesis model can run without a cloud service, directly in a browser. But the demo alone does not prove that the model can replace larger systems: Underdog describes it as a foundation for local use cases such as assistants and personalized podcasts.