AI
Underdog Demonstrates 32MB Offline Voice Model Woof and Opens It for Download
Underdog released a browser game made with the 32MB voice model Woof to showcase offline voice-generation capabilities. The model has 31000000 parameters, can run without an internet connection on most devices and is open for download. The company previously also launched the 167MB Bark speech-to-text model.
Against a backdrop in which large AI models generally rely on heavy computing power, startup Underdog demonstrated Woof, a small voice model aimed at edge devices. The company released a retro-technology-themed browser game set in a room with a typewriter, CRT monitor and desk lamp; users can type anything on the typewriter using the keyboard and the game will read it aloud, or rotate the music tuner to generate different voices.
Underdog said these interactions are completed in the browser by its voice-generation model and can run without an internet connection on most devices. The Woof model is 32MB in size, has 31000000 parameters and is open for download to everyone. The report suggested that although such a small model is unlikely to replace large models, it is sufficient to support local assistant experiences or personalized podcasts.
Ahead of this demonstration, Underdog had weeks earlier launched the 167MB Bark speech-to-text model, also aimed at transcription and dictation on edge devices. Founder Sigil Wen said he wants to build AI systems that run on users’ devices and let users own their data, rather than using note-taking tools such as Granola. He also said the company has released a 27000000000-parameter model that can run on laptops and has a waitlist application positioned as a locally hosted alternative to Muse and Instinct.