Underdog Releases 32MB Woof Speech Model in Browser Demo

Underdog's New Small Speech Model

Underdog, a startup specializing in AI models for edge devices, has unveiled a new speech generation model named Woof. This model is notably small, at only 32MB, and is demonstrated through a retro-themed browser game. The game allows users to type text and have it spoken back, or generate music, all powered by the local model.

Local Operation and Accessibility

The Woof model operates entirely within the browser, requiring no internet connection, and is designed to run on most devices. It features 31 million parameters and is available for public download. This approach contrasts with larger, cloud-based AI models, emphasizing privacy and local data ownership.

Potential Applications and Previous Releases

While not intended to replace larger AI models, Woof could enable local assistant functionalities or personalized audio content. This release follows Underdog's earlier introduction of Bark, a 167MB speech-to-text model also designed for on-device transcription and dictation. The company also offers a 27-billion-parameter model for laptops and a waitlisted app for locally hosted AI.

Founder's Vision for Private AI

Sigil Wen, founder of Underdog, stated his motivation for building these local AI systems stems from a desire for users to own their data, moving away from cloud providers. He expressed a belief that capable AI models will soon run on existing user devices, promoting open, private, and American-made AI.