Underdog's Woof is a tiny 32MB speech model that runs offline

[RogueDepot]

Not every AI model needs to be huge. TechCrunch reports that Underdog has shown off Woof, a text to speech model with about 31 million parameters that takes up roughly 32MB

Woof can generate both speech and music, and it runs in a web browser and on most devices without an internet connection. It is free to download

The demo is a retro style room game where typed text is spoken back to you and a tuner knob generates music. Underdog has previously released a 167MB speech to text model called Bark and a 27 billion parameter model that runs on laptops

Founder Sigil Wen said the company wants to train the best open models that are American made, open and capable, and that allow people to have private AI. The offline claims come from the company, and Woof will not replace large models

A speech model small enough to run anywhere is useful for accessibility tools, games and devices with poor connections. Keeping everything on the device is also good for privacy. It will be interesting to hear how natural it sounds compared with bigger cloud voices

Would you use a small offline voice model in your projects? And do you prefer on-device AI for privacy reasons?

Underdog shows off a 32MB speech model in new demo | TechCrunchtechcrunch.com