Underdog shows off a 32MB speech model in new demo
Underdog has launched Woof, a 32MB speech model capable of generating speech and music directly within a web browser, reportedly without an internet connection. The company demonstrated its capability through a retro-themed game designed to run locally on devices. This follows their earlier 167MB speech-to-text model, Bark, also for edge devices.
Underdog has introduced its 32MB Woof speech model, capable of generating both speech and music within a browser-based demo. This model, along with its 167MB Bark speech-to-text counterpart, is designed to operate directly on a user's device, reportedly without needing an internet connection. This strategy contrasts with the broader industry trend of developing larger AI models that typically demand significant cloud-based processing power.
The company's founder, Sigil Wen, stated his aim to enable users to "own their data" by running AI models locally, framing it as an alternative to "intrusive" cloud providers and services like Granola, Muse, and Instinct. Underdog's approach emphasizes privacy and device-side processing, with its models made available for download and boasting 31 million parameters for Woof.
While these compact models might not replace larger, more powerful AI systems, their local operation could support applications such as personalized podcasts or on-device assistant experiences. The ability to run AI functionalities directly on consumer devices like phones or computers suggests a potential shift towards more private, offline AI capabilities, though the full extent of their performance compared to cloud-based alternatives remains to be seen.
