SeaPRwire – Underdog released a browser game on November 6 that speaks back whatever you type into a virtual typewriter. The voice comes from a model the company calls Woof, which ships at 32 megabytes and, according to Underdog, runs on most devices without a network connection.
The file size is the announcement. Speech synthesis has mostly been sold as a service, where the model sits in a data center and the audio arrives over the wire. Woof is small enough to be downloaded once and executed on the machine in front of you, with 31 million parameters behind it.
Underdog built the demo as a single room: a desk lamp that clicks, a CRT monitor, a typewriter, a music tuner. Type a sentence and the game reads it aloud. Rotate the tuner and it produces tones. The intro warns about “intrusive” cloud providers in the town, which is a marketing line delivered as a joke.
Those interactions are worth taking seriously as a specification, because each one rules something out. Speech has to start generating before the sentence is finished, or the typewriter feels broken. The tuner implies pitch and prosody can be controlled at generation time rather than applied afterward. Neither effect needs a cloud round trip, and the demo is designed to make that visible.
What the demo does not include is any measurement. There is no word error rate, no mean opinion score, no real-time factor, no latency figure for a mid-range phone, and no comparison against the larger models Underdog’s own copy implicitly benchmarks itself against. A visitor can hear that the voice is intelligible. A buyer cannot tell from the page how far the gap to a hosted system runs, or where it widens.
A demo is a claim about a ceiling, not about a floor.
Parameter count gives a rough sense of the trade. At 31 million parameters, the acoustic model, the vocoder, the text front end and the training corpus all weigh more on the output than raw size does, and a mistake in any of them is audible. That is why demos of this kind exist. They show the ceiling a small model reaches on a curated script, where the sentences are short, clean and known in advance.
The browser also constrains what can be shown. Running in a tab means competing for memory with everything else the user has open, and it means inference has to happen fast enough to keep audio streaming. Underdog says the model runs offline on most devices, but does not say which devices failed that test, or what the memory ceiling is.
Power draw belongs in the same list. A voice model running in a tab competes with the display for battery, and on a phone that trade surfaces as heat and throttling.
There is a second model in the family, and its size explains the strategy better than the first. Bark, released a few weeks earlier, is a 167 megabyte speech-to-text model aimed at transcription and dictation. Read together, the two cover the two directions of voice: listening and speaking, both on the device, both downloadable.
Underdog’s founder, Sigil Wen, framed the work around data ownership. He said he started building the app because he did not want to rely on notetakers such as Granola, and preferred a system where the models run on hardware the user controls.
“I think we will be able to run most capable models of today on the devices you already own within months,” he told TechCrunch. “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.”
That last clause is doing more work than the engineering. On-device inference is the only architecture where a privacy claim can be checked by the user, since the audio never leaves the machine. It is also the only architecture where the vendor cannot meter usage, which is why the business model for small models has stayed thin.
The 27 billion parameter claim is the one that needs the most scrutiny. A model that size running on a laptop is possible with aggressive quantization and memory mapping, and the practical result depends on how much of the model stays resident, how fast the storage is, and how much quality the compression costs. Underdog has not published those numbers either.
Wen’s timeline is the part to hold loosely. Saying that most capable models of today will run on hardware people already own “within months” sets an expectation that the company will be measured against, and the gap between a 32 megabyte voice model and a frontier assistant is not a matter of a few quarters of optimization.
Underdog is also selling an app, currently waitlisted, positioned as a locally hosted alternative to Muse and Instinct. That is the commercial shape of the strategy: give away the models as downloads, charge for the experience built on top, and let the privacy argument carry the marketing. It is a shape that has been tried before in other categories of local software, and it tends to produce one of two outcomes. Either the paid layer finds a use that a cloud service cannot serve, usually because the data is too sensitive or the latency too tight, or the free download becomes the whole product and the paid layer never converts. Which of the two happens here depends on whether the app does something the models alone do not, and the announcement does not describe what that is.
The demo does its job as a demo. It puts a voice model in front of people with no install step and no account, and it makes an abstract claim about local inference into something you can hear in ten seconds. What it cannot do is answer the question a platform team would ask before standardizing on it: how does this sound on the sentences we actually need, on the phones our users actually carry, and what happens when the text is not a short line typed into a typewriter.
Those answers will come from evaluation, not from a game. The fact that the game is charming is not evidence either way.