Underdog shows off a 32MB speech model in new demo

Local AI speech models that don't require an internet connection can now run on your device.

Underdog AI, a startup focused on developing AI models for edge devices, has unveiled a new text-to-speech model named Woof. This model is remarkably small, measuring just 32MB, and is demonstrated through a playable browser-based game. The game allows users to type text that the Woof model then speaks, and also generate music through an interactive tuner. Notably, the Woof model operates locally, without requiring an internet connection, positioning it as an alternative to larger, cloud-dependent AI services. This release follows Underdog's earlier launch of Bark, a 167MB speech-to-text model also designed for on-device processing. Underdog's founder, Sigil Wen, aims to create capable, open, and private AI models that can run on existing user hardware, challenging the dominance of cloud-based solutions.

Underdog's demonstration of a 32MB speech model challenges the industry's trend towards massive AI models requiring significant cloud infrastructure. By achieving this small footprint with 31 million parameters, the company signals a viable path for on-device AI, particularly for consumer applications like personal assistants or localized content creation. This approach emphasizes user privacy and data ownership, directly contrasting with the prevalent cloud-based AI services.

The release of Woof and the earlier Bark model indicates a growing market for highly efficient, locally runnable AI. This could affect developers by providing accessible tools for creating offline-capable AI features, and users by offering more private and responsive AI interactions. Future developments will likely focus on increasing the capabilities of these small models and expanding their integration into everyday devices and applications.