Underdog Unveils Compact 32MB Speech Model with Interactive Browser Demo

Image: TechCrunch · Source

Underdog introduces Woof, a 32MB text-to-speech AI model running entirely offline in browsers, demonstrating speech and music generation suitable for edge devices.

In a landscape dominated by increasingly large and resource-intensive AI models, Underdog is taking a different path by focusing on compact, efficient models capable of running directly on consumer devices without cloud dependency. On November 6, 2026, Underdog showcased Woof, a new text-to-speech AI model only 32MB in size, via an interactive browser-based game.

The demo places users in a retro-themed room containing a typewriter and a music tuner. Users can type any phrase, which the game then vocalizes using the Woof model. Additionally, rotating the tuner generates different music sounds, all produced locally within the browser environment. This means Woof runs offline on most devices without requiring internet connectivity, thanks to its 31 million parameters compact architecture.

Underdog emphasizes privacy and user control, positioning Woof as a practical solution for edge devices such as phones and laptops. While this small model cannot match the quality or capabilities of large cloud-based AI systems, it aims to enable local assistant functionalities and personalized audio content creation without data leaving the user’s device.

This release follows Underdog’s recent launch of Bark, a 167MB speech-to-text model also optimized for on-device operation. Company founder Sigil Wen highlighted the goal of creating powerful, American-made open AI models that support privacy by design.

Beyond these speech models, Underdog has developed a larger 27-billion-parameter model designed to run on laptops, and a forthcoming app intended as a locally hosted alternative to cloud AI platforms like Muse and Instinct, currently available via a waitlist. These efforts underscore the startup’s commitment to delivering capable AI experiences directly within user environments, addressing concerns over intrusive cloud-based AI services.

Underdog’s Woof model represents a significant step in making speech AI accessible and private, offering a lightweight alternative for those prioritizing data security and offline functionality in their AI applications.