Underdog announced Woof on November 5: a text-to-speech model that weighs 31.7MB and runs in a browser tab. The demo is charming. The more interesting part is what a model that small does to the cost of putting a voice on a web page.
Underdog's 32MB speech model fits in a browser tab
The first number to hold onto is the file size, because it decides where the model can live. Woof has 31 million parameters and comes in at 31.7MB, and TechCrunch reports that anyone can download it.
That's small enough to ship as a static asset next to your JavaScript. No inference endpoint, no per-character bill, no request that fails when the network drops. Underdog says it runs in a browser and on most devices without an internet connection, which is the kind of claim that only matters if the download is small enough to cache. This one is.
For scale, 31.7MB is roughly the size of a dozen phone photos. A voice model that fits in that budget stops being infrastructure and starts being a file.
What the Woof demo actually puts on screen
The demo is a small browser game, and it's worth describing because it explains the product better than a spec sheet would. You land in a retro-tech room with a typewriter, a CRT monitor, a desk lamp that clicks and a music tuner. Type on the keyboard and the model speaks what you typed.
The game's intro warns about "intrusive" cloud providers in the town, which reads as a jab at hosted speech APIs. Players who defeat a character called Darkcloud get an invite. It's a marketing stunt with a working model inside it, and the working model is the point.
Why a small speech model changes the billing math
Hosted text-to-speech charges by the character, and the meter runs every time someone presses play. That's fine for a demo and awkward for a recipe site that wants to read 400 words aloud to every visitor.
An on-device model moves the cost to the device. You pay once in download size and then in battery, and you stop paying per word. For a site with heavy repeat use, that's a different business model, not a cheaper line item.
The trade is honest. The user's laptop or phone does the work, so a weak device gets slower speech and a hot battery. A cloud model borrows someone else's GPU. Neither is free, and the bill just arrives in a different currency.
Bark, the 27B model and Underdog's waitlisted app
Woof isn't a one-off. Weeks earlier, Underdog released Bark, a 167MB speech-to-text model built for edge devices, for transcription and dictation. The pair covers both directions of a voice interface, and both are sized for hardware people already own.
The company has also released a 27-billion-parameter model that it says can run on laptops, and a waitlisted app it positions as a locally hosted alternative to Muse and Instinct. That's a lot of shipping for a small outfit, and it tells you where the founder thinks the market is going.
Founder Sigil Wen told TechCrunch he started building the app because he wanted his own system, with models that run on users' devices and let them own their data. He also said he thinks most capable models of today will run on devices people already own within months, and that Underdog wants to train the best open, American-made models so people can have private AI.
Treat the timeline as a founder's forecast. The direction is consistent with the products, which is more than you can say for most roadmaps.
What the browser speech model can't prove yet
Can a 31.7MB model sound good enough for a customer-facing voice? Nobody outside Underdog knows yet. There's no word error rate, no naturalness score, no listening test. Underdog hasn't published one for Woof, and no independent evaluation exists. That's the gap between a demo that makes you smile and a model you'd put in front of customers.
A model this small is unlikely to replace the bigger speech systems, as TechCrunch notes. It may be enough for a local assistant or a personalized podcast, where a slightly flat voice is an acceptable trade for privacy and zero marginal cost.
Three limits are worth naming. The claim is "most devices," not every phone, browser or smart speaker. The demo needs a modern browser with JavaScript, and performance depends on the hardware underneath. And the download is open, but the licence terms for commercial use aren't established in the coverage, so check before you build a product on it.
What builders can ship with on-device TTS this month
If you've been putting off voice because the API bill scared you, this is the month to prototype. A pronunciation button on a language site, an article reader that works on a plane, a kiosk demo with no backend: all of those fit inside 31.7MB and a browser tab.
Worth a prototype.
Test the ugly parts before you promise quality. Feed it numbers, acronyms, product names and a paragraph with three commas in one sentence. Try it on the oldest laptop in the office and on a mid-range Android phone. Then read the licence.
Woof proves the size argument. The quality argument is still open, and the first team to publish a listening test will settle more than Underdog's marketing page can.
