Underdog Launches 32MB 'Woof' AI Speech Model That Runs Offline on Everyday Devices

AI startup Underdog has introduced Woof, a text-to-speech model that weighs in at just 31.7MB and is designed to generate spoken audio directly on a user's phone or computer rather than on remote cloud servers. The company has also released a free browser-based game that lets users type text and hear it read aloud, showcasing how the model performs in a lightweight environment.

According to the company, Woof contains approximately 31 million parameters and takes up a little over 31 megabytes of storage. That footprint is far smaller than the large language and speech models that typically depend on powerful server infrastructure, allowing the system to operate on ordinary consumer hardware.

A key feature is offline capability. Underdog says that once the necessary files are stored locally, the model can run on most devices without an internet connection, meaning speech is produced on the device itself instead of text being sent to a remote server. The company has made the model freely available for anyone to download.

Underdog suggests the small size could suit use cases such as offline voice assistants, text-to-speech utilities and personalised audio content. Local processing may also appeal to users who prefer their text to stay on their own hardware rather than travel to an external service.

The company has acknowledged that a compact model does not automatically match the quality or range of much larger AI systems, so Woof is likely to be seen as a lightweight alternative rather than a direct replacement for cloud-based voice platforms. Even so, the release reflects a growing interest in on-device AI that balances privacy, connectivity and hardware constraints.