Underdog Woof review 2026. Honest scoring on voice quality, the 32MB offline model, privacy, the non-commercial license, and who it actually fits.
Underdog Woof is an impressive engineering milestone: a ~32MB, 31M-parameter text-to-speech model that runs entirely on-device and offline, even inside a browser, keeping your text fully private. As a voice for real creator work it is narrow — English-only, modest in realism against cloud TTS, and, most limiting, licensed for non-commercial use only without separate permission. Score it as a strong privacy-and-portability proof of concept and a poor fit for anyone shipping monetized content today.
Underdog released Woof on November 5, 2026 — a text-to-speech model of roughly 32 megabytes with around 31 million parameters that runs on-device, offline, and even entirely inside a web browser. The company demoed it with a retro browser game at underdog.ai/escape where you type on a virtual typewriter and hear the words spoken back, synthesized locally in the tab. The pitch, from founder Sigil Wen, is private AI that never leaves your machine.
I run a content engine, so the disclosure is upfront: Kompozy is generation-and-publishing software, not a TTS model, and Woof is not really a competitor to it — they sit in different categories. That is exactly why this review stays honest about what Woof is genuinely good at. The headline achievement is real: squeezing usable speech synthesis into 32MB that runs offline is hard, and the privacy story is the best in its class.
The caveats are equally real and they decide the fit. Woof is English-language, its quality figures are the company's own rather than independently reproduced, and it is licensed for non-commercial use only without separate permission. For a creator who earns from their work, that license is close to disqualifying on its own. Everything below is scored against the model's state as of 2026-11-07, and against what a creator — not a researcher — needs from a voice.
Woof is an open-weight, on-device text-to-speech model from the startup Underdog. At about 32MB (roughly 31.7 MB for a single-voice build) and ~31 million parameters, it is small enough to download and run on most modern devices, and to execute entirely in a browser tab with no server round-trip. It converts written text into spoken audio locally, which means no internet connection is required once the files are present and no text is sent to a third party. The weights are published for download (hosted on Hugging Face). It is a voice engine and nothing more. Woof does not write scripts, edit audio into a video, generate captions or images, or publish anywhere — it outputs a waveform. It is English-language, its accuracy claims are self-reported, and it is licensed for non-commercial use only without separate permission. Woof is the sibling of Bark, Underdog's roughly 167MB on-device speech-to-text model; Bark transcribes audio to text, Woof speaks text as audio.
Woof fits a developer, researcher, or privacy-focused tinkerer who wants local, offline speech with no cloud dependency and no API bill — for accessibility read-aloud, on-device app features, or non-commercial demos. It is an excellent fit for anyone whose priority is keeping text on their own machine. It is a poor fit for the working creator: the non-commercial license blocks monetized use, the English-only scope and modest realism limit polish, and it generates only audio, so none of the surrounding work — video, captions, publishing — is addressed. If your job is shipping on-brand content you earn from, Woof is the wrong tool category, not just a weaker option.
| Dimension | Score | Why |
|---|---|---|
| Voice quality & naturalness | 3.0 / 5 | Usable and clear for its size, but a 31M-parameter model does not match large cloud TTS on realism, and the quality figures are self-reported. |
| Model size & portability | 4.8 / 5 | About 32MB is remarkably small — it downloads fast and ships inside an app or a web page. |
| Offline & privacy | 4.9 / 5 | Runs entirely on-device with no internet and nothing sent to a server — the strongest privacy story in the category. |
| Ease of setup | 3.5 / 5 | Open weights on Hugging Face plus an in-browser demo lower the bar, but it is still a model to wire in, not a finished app. |
| Language coverage | 2.5 / 5 | English-language; no multilingual support is claimed, which narrows who it serves. |
| Licensing for creators | 1.8 / 5 | Non-commercial use only without separate permission — close to disqualifying for anyone who monetizes content. |
| Commercial production readiness | 2.0 / 5 | Outputs raw audio only; no captions, video, or publishing, and the license blocks paid use. |
| Value (free, open weights) | 4.0 / 5 | Free to download and run with no API cost; the value is capped by the non-commercial license rather than price. |
Woof has no price tag in the usual sense: the weights are free to download and run, and because generation happens on your own device there is no per-character or per-minute cloud bill. In a market where commercial TTS is metered by the character, "free and local forever" is a genuine draw for high-volume or privacy-sensitive use.
The cost is not money, it is the license and the ceiling. Woof is permitted for non-commercial use only without separate permission, so the moment your audio touches a monetized video it is outside the grant — which, for a creator, is the use that matters. On top of that you are responsible for the compute to run it (trivial, given the size) and for everything the model does not do: scripting, editing, captioning, and publishing. So the honest read is that "free" describes the download, not the workflow.
Against a creator's real alternatives the comparison is apples-to-oranges by design. A commercially-licensed voice inside a content engine costs money but can legally ship in paid work and arrives wrapped in video, captions, and distribution; Kompozy Starter, for example, is $199/mo for generation across eighteen formats and publishing to nine destinations, voice included. Woof is free and private but non-commercial and audio-only. Which is "cheaper" depends entirely on whether you are allowed to use the output for what you need.
| Use case | Fit | Why |
|---|---|---|
| Developer adding offline, private read-aloud to an app | Strong | On-device, tiny, and free — exactly what Woof is built for. |
| Privacy-focused user who wants text spoken without the cloud | Strong | Runs entirely locally with nothing sent to a server. |
| Non-commercial demo, research, or hobby project | Strong | The non-commercial license fits this use squarely. |
| Creator drafting and hearing a script before production | OK | Fine as a private scratchpad, but the audio cannot ship in paid work. |
| Voiceover for a monetized video | Weak | The non-commercial license blocks paid use without separate permission. |
| Multilingual narration | Weak | English-language only; no multilingual support is claimed. |
| Turning one idea into finished, published, on-brand content | Weak | Woof outputs audio only — no captions, video, or publishing. |
The cleanest way to place Woof against Kompozy is to say plainly that they are not the same kind of thing, and the review would be dishonest if it pretended otherwise. Woof is a voice model — a single, private, offline component. Kompozy is a content engine that generates finished video, carousels, quote cards, blogs, and newsletters and publishes them across nine destinations on a schedule. You would not choose "Woof or Kompozy" any more than you would choose "a microphone or a studio."
Where they actually touch is the voice, and that is where the license decides it. If you earn from your content, Woof's non-commercial grant takes it off the table for the one job a creator needs a voice for; Kompozy's Persona Shorts carry a commercially-usable native voiceover plus captions, inside a pipeline that also builds and ships everything around the voice. So the honest positioning is not that Kompozy beats Woof on speech synthesis — it may not — but that Kompozy solves the creator's whole problem commercially, while Woof solves one slice of it privately and non-commercially.
For what it sets out to do — tiny, private, offline speech synthesis — it is genuinely impressive, fitting usable text-to-speech into about 32MB that runs locally with no cloud. For polished, multilingual, commercially-usable voice it is limited: English-only, modest realism against cloud TTS, and licensed for non-commercial use only.
The weights are free to download and run, and because generation happens on your device there is no per-character cloud bill. The real cost is the license, not the price: Woof permits non-commercial use only without separate permission.
Not as it stands. The non-commercial license means using Woof's voice in monetized content requires separate permission from Underdog. For paid videos you need a commercially-licensed voice, such as the native text-to-speech in Kompozy's Persona Shorts.
They optimize for opposite things. ElevenLabs wins on voice realism, multilingual range, and a commercial license; Woof wins on size, offline operation, and privacy. Pick ElevenLabs for polished commercial voiceover, Woof for private, local, non-commercial speech.
Yes. Once the model files are local, Woof generates speech on-device with no internet connection, and the browser-game demo runs the synthesis entirely in the tab. No text is sent to a server — that privacy is the model's main selling point.
Woof outputs audio and nothing else. It does not write scripts, edit video, generate captions or images, or publish to any platform. Turning a narration idea into finished, published content is the job of a content engine like Kompozy, not a standalone voice model.
No, they are opposites. Bark is Underdog's roughly 167MB on-device speech-to-text model that transcribes audio into text; Woof is the ~32MB text-to-speech model that speaks text as audio.
Developers, researchers, and privacy-focused users who want local, offline, free speech synthesis for accessibility, app features, or non-commercial projects. Creators shipping monetized content are better served by a commercially-licensed voice inside a publishing workflow.