AI startup Underdog has released the Underdog Woof AI voice model, a text-to-speech system that is just 32MB in size and can generate speech directly on a phone or computer, with no need for cloud servers.
The company has also launched a free browser-based game to show what the model can do. Users type text and hear it spoken aloud, with every voice in the game generated by Woof itself.
ProPakistani reported that the model is available for anyone to download, and that it runs locally once the files are on the device.
How the Underdog Woof AI voice model works
Woof contains about 31 million parameters and takes up just 31.7MB of storage. That is tiny by the standards of modern AI. Most voice models are hundreds of times larger and run on powerful servers in data centres.
According to Underdog, Woof can run on most devices without an internet connection once the necessary files are available locally. Speech is generated on the user’s own device, rather than sending text to a remote server and waiting for audio to come back.
That design brings two clear benefits. The first is privacy: the text never leaves the device. The second is independence: the tool keeps working when the connection drops.
The trade-off is quality. The company is open about the fact that a model this small cannot match the range or the polish of much larger AI systems. For short, clear speech it does the job; for long narration the bigger models still lead.
A game that demonstrates the technology
To show off Woof, Underdog built a free browser game called Escape. It takes place in a retro-themed room with a typewriter, an old CRT monitor, a desk lamp and a music tuner.
Players type messages on the virtual typewriter, and Woof turns the text into spoken audio. The game also lets users interact with other objects, including a music tuner that produces different sounds and music.
The game carries a pointed message. Its introduction refers to intrusive cloud providers, which reflects Underdog’s focus on running AI models locally instead of depending on remote servers.
Underdog announced the model on X on November 5, writing: “We made a game for our new frontier text-to-speech model, Woof. 31.7 MB, so small that it runs in your browser.” Players who defeat the character Darkcloud in the game receive an invite.
Part of a wider offline AI push
Woof follows the company’s earlier release of Bark, a 167MB speech-to-text model built for transcription and dictation on local devices. While Woof turns written text into speech, Bark does the opposite, turning spoken words into text.
Underdog has also introduced a 27-billion-parameter AI model designed to run on laptops. It is developing a separate application that will let users interact with locally hosted AI models as an alternative to cloud-based assistants.
Founder Sigil Wen told TechCrunch that he started building these tools because he wanted an AI system that could run on his own devices and give users greater control over their information. The company’s goal is to reduce dependence on cloud computing while making capable AI tools available on hardware people already own.
Why offline AI matters in Pakistan
The idea of AI that works without the internet is easy to dismiss in places where connectivity is fast and cheap. In Pakistan, the case is stronger.
Mobile data costs money, and connections drop, especially outside the big cities. A voice tool that runs on the phone itself keeps working on a bus, in a village, or during a power cut, as long as the battery holds.
That is also why the government is paying attention to how AI reaches the rest of the country. Officials have pointed to cost, connectivity and skills as the barriers holding AI back in rural Pakistan, and small, offline models address at least the first two of those problems directly.
There is a security angle too. The National Cyber Emergency Response Team has warned banks and utilities about AI-enabled cyber threats, and tools that keep data on the device reduce one kind of exposure, even if they do not solve the wider problem.
Woof itself is a narrow tool. It generates speech, and nothing else. But it points to a direction of travel: AI that lives on the device, not in a distant data centre. WhatsApp’s new Events feature, which now works for people who do not even use the app, shows the same instinct in a different form. WhatsApp has also let people join calls without an account.
Frequently asked questions
What is Underdog’s Woof?
Woof is a text-to-speech AI model from the startup Underdog. It is 31.7MB in size, with about 31 million parameters, and it generates speech directly on a phone or computer without an internet connection.
How can I try Woof?
Underdog has released the model for download, and has also built a free browser game called Escape that demonstrates it. Players type text on a virtual typewriter and hear it spoken aloud by Woof.
Does Woof sound as good as larger AI voice models?
No. Underdog says the model’s small size means it cannot match the quality or range of much larger systems. It is designed for short, clear speech on everyday devices, not for long-form narration.
What else has Underdog built?
Its earlier release, Bark, is a 167MB speech-to-text model for transcription and dictation on local devices. The company has also built a 27-billion-parameter model for laptops and is developing an app for locally hosted AI models.
