Underdog has released a 32MB speech model that runs entirely in a web browser without an internet connection.
The demonstration is a game set in a room with retro tech equipment, including a typewriter and a CRT monitor. Clicking a desk lamp switch triggers a sound, while the intro warns users about “intrusive” cloud providers in the town. This serves as a critique of cloud-based models.
Players interact with the typewriter and a music tuner. Typing text on the keyboard causes the game to speak it back. Rotating the tuner generates different kinds of music.
Underdog’s speech generation model uses 31 million parameters. The company states it runs on most devices offline and is open for everyone to download.
While this small model will likely not replace larger systems, it may suffice for local assistant tasks or creating personalised podcasts.
This demo follows the launch of Bark, a 167 MB speech-to-text model also designed for edge devices to handle transcription and dictation.
Sigil Wen, the founder, told TechCrunch he built the app because he wanted a system that runs on users’ devices rather than using note-taking tools like Granola. He aims to let people own their data.
“I think we will be able to run most capable models of today on the devices you already own within months. We just want to train the best open models that are American-made, that are open, that are capable and allow people to have private AI,” he said.
Underdog has also released a 27-billion-parameter model for laptops and a waitlisted app positioned as a locally-hosted alternative to Muse and Instinct.
What it means
The project demonstrates that voice interaction does not require massive servers. Users can own the software and keep their data private on their own hardware.
