Underdog shows a browser game running its 32MB Woof speech model

Underdog, an AI startup, released a browser-based game built around Woof, a text-to-speech model the company says is about 32 megabytes in size with 31 million parameters. A text-to-speech model simply converts written text into spoken audio. Once loaded, Woof can run on most devices entirely offline, with no internet connection needed, since the whole model lives on the device rather than on a remote server. In the demo, typed text is spoken back aloud directly in the browser, and a built-in tuner can generate music, all processed locally on the user's device. Underdog has made the model's weights, meaning the trained parameters that define how it behaves, available for anyone to download. This follows the company's earlier release of Bark, a larger 167 megabyte speech-to-text model also designed to run directly on-device.

Why it matters

Small, fully local speech models that run in a browser point toward voice features that work without sending audio to the cloud, which matters for privacy and offline use.

Source: TechCrunch