Underdog shows off a 32MB speech model in new demo

Underdog, a startup building small AI models that run on a user's own device, has released a browser game to show off a new text-to-speech model called Woof that is only about 32MB in size, TechCrunch reports.

The game is set in a retro-styled room with a typewriter, a CRT monitor and a music tuner. Players can type anything on the typewriter and hear it spoken back, or turn the tuner to generate different kinds of music. Underdog says all of it is produced in the browser by Woof, which has 31 million parameters, can run on most devices without an internet connection and is free for anyone to download.

TechCrunch notes that a model this small is unlikely to replace larger ones, but could be enough for local assistants or personalized podcasts. The demo follows Bark, a 167MB speech-to-text model for transcription and dictation that Underdog launched a few weeks earlier. The company has also released a 27-billion-parameter model that runs on laptops, and has a waitlisted app it pitches as a locally hosted alternative to assistants such as Muse and Instinct.

Founder Sigil Wen told TechCrunch he wants to train open, American-made models that let people keep their data on their own machines, and that he expects today's most capable models to run on devices people already own "within months".