Underdog's new voice model fits in 32MB. It still beats a 41GB rival on hard sentences
Underdog's Woof text-to-speech model downloads in under 32MB and runs offline in a browser, and the company says its two-voice build reads numbers and dates more accurately than Fish Audio's 41.5GB Drama 3.
Underdog shipped a voice model small enough to fit on a floppy disk, if anyone still had one lying around. Woof, the company’s new text-to-speech model, downloads in 31.7MB for one voice or 40.6MB for two, and once it’s loaded in your browser it keeps working with the Wi-Fi off. Nothing you type leaves the machine.
To show it off, Underdog built a tiny browser game instead of a spec sheet. You land in a dark room with a typewriter, a radio and a desk lamp. Type anything on the typewriter and the room speaks it back in real time. Rotate a tuner knob and it hums different kinds of music. The intro jokes about an “intrusive” cloud utility stalking the town, which is Underdog’s way of saying: none of this called a server.
It’s a cute wrapper. The model underneath is the actual story.
How small is small
Underdog’s own numbers, published on its site, put Woof’s one-voice build at 30.7 million parameters and 31.7MB. The two-voice version is 38.2 million parameters and 40.6MB. For comparison, Fish Audio’s Drama 3 (the model Underdog picked as its benchmark) runs 35.9 billion parameters and a 41.5GB download, more than a thousand times larger than either Woof build.
| Model | Download | Parameters | Hard-sentence error rate |
|---|---|---|---|
| Woof, one voice | 31.7 MB | 30.7M | 2.06% |
| Woof, two voices | 40.6 MB | 38.2M | 1.42% |
| Fish Audio Drama 3 | 41.5 GB | 35.9B | 1.89% |
On a 500-sentence stress test packed with phone numbers, prices, dates and codes, Underdog says two-voice Woof gets 1.42% of words wrong, against 1.89% for Drama 3. The one-voice model lands at 2.06%, roughly the same error rate as the model 1,300 times its size. Speed-wise, Underdog clocked one-voice Woof turning 13.2 seconds of speech into 0.73 seconds of compute on an Apple M5 Max CPU thread, using plain PyTorch with int8 weights.
My read: those are good numbers, and the size claim alone is the headline. But every figure above comes from Underdog’s own benchmark page, measured against a model it chose to compare against, on hardware it chose. Nobody outside the company has run Woof against Drama 3 yet. Call it promising, not proven.
Why the sentence-reading problem is real
Small voice models usually trip on exactly what Underdog tested: numbers, dates and codes read in ways no human would say them. A phone number like 020 7946 0018 can come out as one enormous nine-digit figure instead of a string of individual numbers. Underdog built a text normalizer that rewrites that kind of input before the model ever sees it, turning “$1,249.50” into “one thousand two hundred and forty nine dollars fifty cents” and “14/03” into “the fourteenth of March.” That’s a real, if unglamorous, problem for anything that reads order confirmations, boarding passes or appointment reminders out loud, and it’s the part of the demo that has nothing to do with the room-escape game.
Not Underdog’s first small model
This is Underdog’s second edge-focused release in a matter of weeks. In October the company put out Bark, a 167MB speech-to-text model built for on-device transcription and dictation, and it’s also released a 27-billion-parameter model sized for running on a laptop. Underdog founder Sigil Wen told TechCrunch he started the company after getting frustrated with cloud notetakers like Granola and wanting a system that keeps a user’s audio on their own device. “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.
Woof is open weights, but not Apache 2.0: Underdog licenses it under CC BY-NC 4.0, non-commercial use only. That’s the same kind of two-tier licensing fight that showed up when Gradium shipped its Phonon text-to-speech model last week, where the better-sounding of its two sizes also carried a non-commercial license. If you want to actually ship Woof inside a paid product, read the license before you get attached to the demo.
My read on where this fits
A 32MB model that runs in a browser tab isn’t going to replace a studio voice actor or a flagship cloud TTS system, and Underdog isn’t claiming it will. What it’s good for is the boring, constant stuff: an app that reads back a confirmation number, an assistant that works on a plane with no signal, a personalized podcast feed generated entirely on-device. That’s a narrower claim than “best voice model,” and it’s also one Underdog can actually back up with a download link instead of a press release.
Related: Phonon TTS tops a new speech leaderboard. The better model isn’t the free one · Meta’s new model costs a third of Claude Haiku’s price and still beats it on agent tasks · Tether built an AI model that fits on your phone and still beats a bigger one
Sources
Note on sources: the size, speed and accuracy numbers above are Underdog’s own, published on its Woof page and the Hugging Face model card; TechCrunch’s reporting confirms the demo, the Bark model and Sigil Wen’s quotes. No outside lab has re-run Underdog’s benchmarks yet.
- Underdog shows off a 32MB speech model in new demo - TechCrunch, November 6, 2026
- Underdog Woof - Underdog’s own benchmark and specs page
- Underdog-Woof model card - Hugging Face