The startup Underdog shipped Woof on November 5, 2026 — a roughly 32MB, 31-million-parameter speech model that turns text into audio on-device, with no cloud and no internet, demoed inside a retro browser game.
2026-11-07 · by Moe Ameen
Underdog, a small AI startup, released Woof on November 5, 2026 — a text-to-speech model of roughly 32 megabytes (about 31.7 MB for a single-voice build) with around 31 million parameters. First reported by TechCrunch, the model turns written text into spoken audio entirely on the device: once the files are local, generation needs no internet connection and nothing is sent to a server. Founder Sigil Wen framed the release around private, on-device AI that keeps a user's data on their own machine.
To show it off, Underdog shipped a browser game at underdog.ai/escape — a retro-styled virtual room where you type on a typewriter and hear Woof speak the words back, generated locally in the tab, alongside other objects like a CRT monitor and a music tuner. The demo's intro screen takes a pointed shot at "intrusive" cloud voice providers, which is the whole pitch: the speech never leaves your browser.
Underdog published the weights for download (hosted on Hugging Face), but the license permits non-commercial use only without separate permission — the single most important caveat for anyone who earns from their content. The model is English-language, and the quality figures are the company's own rather than independently reproduced; a model this compact will not match large cloud TTS systems on voice realism. Woof follows Bark, a roughly 167 MB on-device speech-to-text model Underdog shipped weeks earlier for local transcription — the opposite direction, turning audio into text.
Woof is a genuine milestone for private, offline voice — but a raw audio clip is the smallest piece of a content workflow. The job that eats a creator's time is turning a script into a finished, captioned, on-brand video and getting it in front of an audience on every platform. That is what [Kompozy](/) is built for, and for narrated content it already carries its own commercially-usable voiceover: [Persona Shorts](/glossary/persona-shorts) render a talking-head avatar with native text-to-speech plus auto-burned captions, and the faceless video formats narrate over stock footage — none of it under a non-commercial license.
So the practical move today is to treat Woof as a private scratchpad for hearing a script read aloud, and Kompozy as the production line. From one idea, Kompozy generates the clip, brand-exact [carousels](/glossary/hyperframes), quote cards, per-platform text, a blog article, and an email newsletter, all held to one [Persona Brief](/glossary/persona-brief), then [Autopilot](/glossary/autopilot) schedules and publishes the set across eight social platforms plus blog and email behind a per-post review gate. For the full workflow see our [Underdog Woof tool guide](/ai-tools/underdog-woof); for the honest trade-offs, the [Woof review](/reviews/underdog-woof) and the [Woof alternative](/alternatives/underdog-woof) pages.
Woof is a text-to-speech model from the startup Underdog, released November 5, 2026. It is about 32MB with roughly 31 million parameters and converts written text into spoken audio entirely on-device — it can run offline, including inside a web browser, with nothing sent to a server.
Not without separate permission. Underdog published Woof under a license that allows non-commercial use only, so using its voice in content you earn from is not permitted by default. For monetized work you need a commercially-licensed voice — for example the native text-to-speech built into Kompozy's Persona Shorts.
No. Once the model files are downloaded locally, Woof generates speech on your device with no internet connection — the browser-game demo runs the synthesis entirely in the tab. That on-device design is the point: your text never leaves your machine.
For voice realism, no — a ~32MB model will not match large cloud systems like ElevenLabs, and its quality figures are self-reported. Woof wins on privacy, offline operation, and size. They solve different problems: Woof is a private, local voice; ElevenLabs is a high-fidelity commercial one.