Startup Underdog shows off a speech model of just 32 MB in new demo
US startup Underdog has released Woof, a game built on its new speech model with 31 million parameters that runs in the browser without internet. The model is open for download.
Underdog, a startup building small artificial intelligence models meant to run directly on users' own devices, has released a demo game called Woof. It is powered by the company's new voice model, which is only 32 MB in size. The release was reported by technology outlet TechCrunch.
Most AI companies are currently racing to build ever larger models that demand huge amounts of computing power. Some researchers and startups are taking the opposite approach, creating compact models that can run on edge devices such as phones and computers.
A game that runs in the browser
Woof is set in a room with a retro tech theme, featuring a typewriter, a CRT monitor and a desk lamp that makes a sound when its switch is clicked. The intro warns players about "intrusive" cloud providers, which appears to be a jab at cloud-based models.
Players can type anything on the typewriter with their keyboard, and the game speaks the text back. Turning a tuner generates different kinds of music. Everything runs in the browser using Underdog's speech generation model, which the company says can work on most devices with no internet connection. The model has 31 million parameters and anyone can download it.
A model this small is unlikely to replace larger ones, but it could be enough to power local assistant experiences or personalized podcasts.
Earlier releases
The demo arrives weeks after Underdog launched Bark, a 167 MB speech-to-text model also built for edge devices, aimed at transcription and dictation. Founder Sigil Wen told TechCrunch he began building the app because, rather than using note-taking tools such as Granola, he wanted his own system running on users' devices and letting them own their data.
Wen expects that within months most of today's most capable models will run on devices people already own. The company's aim is to train the best open, American-made models so that people can have private AI.
Beyond the two voice models, Underdog has also released a 27-billion-parameter model that can run on laptops, plus a waitlisted app that it positions as a locally hosted alternative to Muse and Instinct.
