Underdog Woof: 32MB speech AI model runs offline in a browser
By Anika SharmaAI & Innovations Editor, Peepals Tech
Representative image: theglassdesk / Pixabay
US start-up Underdog has released Woof, a 31.7 MB open-weight English text-to-speech model that it says runs offline on most devices, reported 6 November 2026.
What Happened
Underdog, a US AI start-up, has released Woof, a text-to-speech model of about 31.7 megabytes that it says can run on most devices without an internet connection. The company showed it off in a small browser game in which text typed by the player is read aloud by the model running inside the browser. Technology news sites reported the release on Friday, 6 November 2026, and the model's weights, the trained numbers that make it work, have been made available for download under a non-commercial licence.
Key Facts
- Model: Woof, a text-to-speech (TTS) model, which turns written text into spoken audio.
- Size: about 31.7 MB, with roughly 31 million parameters, according to Underdog; it is often rounded to 32 MB.
- Voices and language: two English voices, named Juno and Mira; reports do not mention any Indian language.
- Licence: open weights under Creative Commons CC BY-NC 4.0, which allows sharing and adaptation but not commercial use.
- Demo: a retro-themed browser game with a typewriter and an old-style CRT monitor; once loaded, it works without Wi-Fi, the company says.
- Claim: Underdog says Woof beats models 1,000 times its size on word error rate, a measure of how often words come out wrong; this is the company's own test.
- Founder Sigil Wen has said the aim is to train strong American-made open models that people can run privately on devices they already own, according to reports.
Why It Matters
Most voice features in apps today send text to a company's servers, where a large model produces the audio and sends it back. That needs a steady internet connection, costs the app maker money for every second of speech, and means the text leaves the user's device. A model small enough to sit inside a web page changes all three: it can speak offline, it costs nothing per use once downloaded, and the text never has to leave the phone or laptop.
The size is the striking part. A 31.7 MB file is smaller than many phone games and many photo-heavy web pages. If such models are good enough, developers could add read-aloud buttons, simple voice assistants or spoken lessons to apps that work in places with weak signals. Reviewers have been more careful than the company: a model this small is unlikely to replace large speech systems, though it could be enough for local assistant features or personalised audio such as podcasts.
For Indian readers, the limits matter as much as the promise. Woof speaks English only, with two voices, and its licence bars commercial use, so Indian app makers cannot simply build it into paid products. The broader trend, however, is directly relevant to India, where many users rely on budget phones and patchy mobile data. Small on-device speech models in Indian languages would let reading aids, farm advisories and government service apps talk to users without a data connection. Several Indian teams already work on speech models for Indian languages, and releases like Woof show how small such models can be made.
Peepals calculation: Underdog says Woof beats models 1,000 times its size. With about 31 million parameters, a model 1,000 times larger would have about 31 billion parameters (31 million × 1,000 = 31,000 million). In file terms, 1,000 copies of a 31.7 MB model would come to about 31.7 GB (31.7 × 1,000 = 31,700 MB), far more than most phones can spare for one feature. Our calculation from the reported figures.
| Feature | Woof (as reported) |
|---|---|
| Type | Text-to-speech |
| Size | About 31.7 MB, about 31 million parameters |
| Languages and voices | English; two voices, Juno and Mira |
| Works offline | Yes, once loaded, the company says |
| Licence | CC BY-NC 4.0 (non-commercial) |
Impact
Short-term: Developers and hobbyists can try the model in the browser demo and download the weights for non-commercial projects. Independent tests of the company's accuracy claim are likely to follow.
Long-term: If tiny models keep improving, more voice features may move from cloud servers onto phones and laptops, cutting costs for app makers and keeping users' text private. That shift would matter most in markets with weak connectivity, including much of rural India.
Who is affected: App and web developers, accessibility tool makers, students and researchers working on speech, users with low-end devices or poor internet, and companies that sell cloud-based voice services.
Key Takeaway
Underdog's 31.7 MB Woof model shows that a usable English text-to-speech system can now fit inside a web page and run offline, though its non-commercial licence and English-only voices limit what Indian developers can do with it today.
Questions and Answers
What is Underdog's Woof model?
Woof is a text-to-speech model released by the US start-up Underdog, reported on 6 November 2026. It is about 31.7 MB in size, has roughly 31 million parameters and, according to the company, runs on most devices without the internet.
Can Woof speak Hindi or other Indian languages?
No language other than English has been reported. Underdog describes Woof as an English model with two voices, Juno and Mira.
Can Indian companies use Woof in their apps?
Not in commercial products under its current licence. The weights are released under CC BY-NC 4.0, which allows non-commercial use, sharing and adaptation, but not commercial use.
Is Woof really better than much larger models?
That is Underdog's claim, based on its own word error rate tests. Reviewers have said a model this small is unlikely to replace large speech systems, though it could serve local assistant features.
Disclaimer: Prepared by the Peepals newsroom from publicly available sources with AI assistance. Information is accurate to the best of our knowledge at the time of publication and may change. Images may be representative. Not professional advice. Report an error via our contact page.
By Anika Sharma, AI & Innovations EditorCovers AI & Innovation and more for Peepals Tech. All stories by Anika Sharma
H-1B 60-day grace period: comments on DHS plan to end it close today
Big Walk AI usage: House House admits LLM help after IGF jury stand
Get the day's stories in your inbox
Peepals Weekly: the week decoded, free. One email a week, unsubscribe in one click.
Anika Sharma is a pen name of the Peepals Global Editorial Team, which researches, fact-checks and publishes every story. The portrait is an illustration.
