Underdog launches a 32MB voice AI model that runs offline in your browser
Underdog's new Woof text-to-speech model is so small it runs locally on your device without internet, proving voice AI can skip the cloud.
Models · Source: TechCrunch
What happened
Underdog just released Woof. It is a new text-to-speech model that fits in a tiny 31.7 MB file. You can run it directly in your web browser using WebGPU technology. Once the model loads, you can turn off your Wi-Fi completely. The software keeps generating voices entirely on your device without sending a single byte to a data center.
The startup showed off the tech with a retro browser game called Escape. You type on a virtual typewriter, and the game speaks the words back to you. It even generates music locally. If you defeat the game's villain, Darkcloud, you earn an invite to Underdog's private desktop assistant beta.
Founder Sigil Wen wants to build capable open models that run locally. His goal is private AI where users actually own their data. Underdog previously released a 167 MB transcription model called Bark and a 27-billion-parameter model for laptops. They are positioning their upcoming app as a locally hosted alternative to tools like Muse and Instinct.
Key facts
- 31.7 MB — Size of the Woof text-to-speech model download
- 18 times — Speed of speech generation compared to real-time on a laptop CPU
- 1.42% — Word error rate claimed by Underdog
- CC BY-NC 4.0 — License for the open-weight model, restricting commercial use
Why it matters
Voice AI usually means massive cloud bills and high latency. Woof flips that dynamic completely. You can now build text-to-speech features into edge devices, mobile apps, or desktop software without paying endless API fees. It makes voice generation viable for applications used in airplanes, remote areas, or anywhere with spotty internet connections. Developers can finally build fast voice tools without relying on massive server farms.
Privacy becomes a default feature, not just a marketing promise. When data never leaves the user device, you can safely build voice apps for highly regulated industries like healthcare or finance. Big cloud providers will face real pressure as developers realize they do not need massive data centers for basic AI tasks. The edge computing trend is accelerating, and the companies selling expensive cloud compute will eventually lose their grip on simple AI workloads.
For builders
Free local voice generation
Woof can generate speech at 18 times real-time speed on a single laptop CPU thread. Developers save massive API costs, while cloud providers lose out on basic text-to-speech revenue. Note that commercial use requires a separate license from Underdog since the open weights use a non-commercial license.
Total data privacy for users
Because the model runs fully offline, user text never goes to a server. You can build private note-taking or reading apps for sensitive data. Companies relying on harvesting user voice data will lose their edge as consumers demand local processing.
My take
Big tech wants you hooked on their expensive cloud APIs. Underdog proves you can put capable voice AI right in the user's pocket. Stop renting servers for simple tasks and start building local.
Original reporting: TechCrunch. This is my rewrite and opinion.