Underdog unveils ultra-compact speech model capable of local device operation

The startup Underdog has introduced Woof, a remarkably small speech model measuring just 32MB, demonstrated through a retro-themed browser game. This model, which possesses 31 million parameters, is designed to function entirely offline on most devices. The company also recently released Bark, a 167MB speech-to-text model, both aimed at enabling edge device functionality. Sigil Wen, the founder of Underdog, stated his motivation was to create systems where users maintain control over their data, contrasting with reliance on external notetakers. Wen expressed confidence that most advanced models could run on existing user hardware within months, emphasizing the goal of developing open, American-made, and private AI solutions. Beyond these voice tools, Underdog has also released a 27-billion-parameter model suitable for laptops and is offering a waitlisted application positioned as a local alternative to services like Muse and Instinct.