A roughly 32MB, 31-million-parameter text-to-speech model from the startup Underdog that generates speech on-device — offline, including entirely inside a web browser.
Last verified · 2026-11-07 · by Moe Ameen
Underdog Woof is a text-to-speech model released by the startup Underdog on November 5, 2026. It is tiny by modern standards — about 32 megabytes (roughly 31.7 MB for a single-voice build) with around 31 million parameters — and its defining trait is that it runs entirely on-device. Once the model files are local, it converts written text into spoken audio with no internet connection and nothing sent to a server. Founder Sigil Wen pitches it as private, on-device AI; Underdog demoed it with a browser game at underdog.ai/escape where you type on a virtual typewriter and hear Woof speak the words, synthesized locally in the tab.
The weights are published for download (hosted on Hugging Face), so developers and tinkerers can run it on most modern devices. Two boundaries matter before you build on it. First, the license: Woof is offered for non-commercial use only without separate permission, which is the critical limit for anyone who monetizes their content. Second, the quality ceiling: it is an English-language model, its accuracy figures are the company's own rather than independently reproduced, and a model this compact does not match large cloud TTS systems like ElevenLabs on voice realism. Woof is the follow-up to Bark, Underdog's roughly 167 MB on-device speech-to-text model for local transcription — Bark turns audio into text, Woof turns text into audio.
Read Woof for what it is: a privacy-first, offline, open-weight narrator. It is a voice engine, not a content tool. It writes no script, edits no video, generates no captions or images, and publishes nowhere. It outputs an audio waveform and stops there.
The natural place for Woof in a real workflow is early and private: before you spend credits or hit a cloud API, draft your script and hear it read aloud offline, for free, with nothing leaving your laptop. That is a genuinely useful prototyping loop. The problem is that it ends at a raw WAV under a non-commercial license — and a WAV is nowhere near a published post. [Kompozy](/) is the layer that takes the idea you validated in Woof and turns it into finished, on-brand, commercially-usable content across every platform.
Here is the concrete hand-off, and it is deliberately not "import the audio." Because Woof's voice is non-commercial and English-only, a monetizing creator does not ship it in a paid video; instead they carry the script forward into Kompozy, which generates the narrated video itself. [Persona Shorts](/glossary/persona-shorts) produce a talking-head avatar with native, licensed text-to-speech and auto-burned captions; the faceless Listicle and Naturalistic formats narrate over stock footage; and from the same source Kompozy also builds brand-exact [Carousels](/glossary/hyperframes), quote cards, per-platform text, a blog article, and an email newsletter — all governed by one [Persona Brief](/glossary/persona-brief) so a week of output sounds like one voice. Then [Autopilot](/glossary/autopilot) schedules and publishes the whole set across eight social platforms plus blog and email, behind a per-post review gate. Woof is the private sketch; Kompozy is the studio and the distribution. For the honest trade-offs, see the [Woof review](/reviews/underdog-woof) and the [Woof alternative](/alternatives/underdog-woof).
Woof is a ~32MB, roughly 31-million-parameter text-to-speech model from the startup Underdog, released November 5, 2026. It converts written text into spoken audio entirely on-device — it runs offline, including inside a web browser, with nothing sent to a server.
Not by default. Underdog licenses Woof for non-commercial use only without separate permission, so you cannot legally use its voice in monetized videos as it stands. For paid content you need a commercially-licensed voice, such as the native text-to-speech built into Kompozy's Persona Shorts.
Yes. Once the model files are downloaded, Woof generates speech on your device with no internet connection — the browser-game demo runs the synthesis entirely in the tab. On-device, private operation is the model's whole point.
Woof only outputs audio, so it does not publish anything. Use it to draft and validate your script privately, then generate the finished, captioned video in Kompozy — a Persona Short with a licensed avatar voice or a faceless clip over stock footage — and let Kompozy schedule and publish it across platforms.