Startup Underdog has released a 32MB speech generation model called Woof, demonstrated through a browser-based game featuring a retro-themed room. The model contains 31 million parameters and is designed to run on most devices without an internet connection. Users can interact with a virtual typewriter to hear text spoken back or use a music tuner to generate different sounds.
Founder Sigil Wen stated that the development of the app was driven by a desire to build systems that run on user devices, allowing individuals to own their data rather than relying on third-party notetakers. Wen expressed a goal of training open, capable models that enable private AI. "I think we will be able to run most capable models of today on the devices you already own within months," Wen said.
The release of Woof follows the recent launch of Bark, a 167 MB speech-to-text model intended for local transcription and dictation. While the company noted that such small models may not replace larger systems, they could be sufficient for driving local assistant experiences or creating personalized podcasts.
Underdog has also released a 27-billion-parameter model capable of running on laptops. The company is currently maintaining a waitlist for an app positioned as a locally-hosted alternative to Muse and Instinct.
