Underdog demos a 32MB voice model that runs in a web browser

In short: Startup Underdog shared a demo of “Woof,” a 32MB text to speech model that can run in your web browser and can also make simple music.

What happened

Underdog released a small browser game to show off its new voice model, called Woof. A voice model is software that turns typed words into spoken audio, like having a computer read a message out loud.

In the game, you see a retro-style room with objects like a typewriter and a music tuner. You can type anything, and the game speaks it back using Woof. You can also turn the tuner to generate different kinds of music.

Underdog says the model is about 31.7MB, small enough to run on most devices. It says it can work without an internet connection because it runs on the device instead of sending your text to a company server (like doing the work in your kitchen instead of mailing it to a factory).

The company also said Woof has 31 million parameters, meaning 31 million adjustable settings that help it produce speech. Underdog is making the model available for anyone to download.

Why it matters

Many AI voice systems depend on cloud servers, which can raise privacy concerns since your text or audio may leave your device. Smaller models that run locally could make it easier to build voice features that keep data on your phone or computer, even if they are not as capable as much larger systems.

Underdog’s demo comes a few weeks after it launched Bark, a 167MB speech to text model designed for transcription and dictation on devices. Founder Sigil Wen told TechCrunch the goal is to let people own their data while still using AI tools.

Source: TechCrunch AI