The feature is initially available by invitation to a limited group of U.S. customers, with a wider launch planned in the coming weeks.
Underdog releases browser game showcasing 32MB Woof speech model
The startup says its downloadable, 31 million-parameter model can generate speech on most devices without an internet connection.

Underdog has introduced a browser-based game demonstrating Woof, its 32MB voice-generation model. The startup says the model can operate offline on most devices and is available for anyone to download.
Players explore a room decorated with retro technology, including a typewriter and a CRT monitor. Text entered on the typewriter is read aloud, while adjusting a music tuner produces different kinds of music. Woof has 31 million parameters.
The demonstration follows the launch weeks earlier of Bark, Underdog’s 167MB model for converting speech into text. Bark is intended for on-device transcription and dictation.
Founder Sigil Wen said he began building the app because he wanted his own alternative to notetaking tools such as Granola, using models that operate on users’ hardware and allow them to retain ownership of their data.
Underdog has also made available a model with 27 billion parameters that can operate on laptops. A separate app has a waitlist, with the company presenting it as an alternative to Muse and Instinct that runs locally.
