A 32 MB voice AI brings local speech to the browser
Woof
Underdog’s small AI model for turning written words into spoken audio.
text-to-speech
A system that turns written words into spoken audio.
parameters
Numbers changed while the AI learns.
What happened
Underdog, a startup building AI for personal devices, has shown a game using Woof, a text-to-speech model. TechCrunch, a technology news site, reported that Woof is 31.7 MB, usually rounded to 32 MB. In the game, users type on a virtual typewriter, and Woof reads the words aloud. A music tuner can create different kinds of music. Underdog says Woof runs in a browser on most devices without an internet connection. The company also says anyone can download it.
The background
AI companies are competing to build larger and more capable models. Larger models usually need more computing power. Other teams are moving in the opposite direction. They are making smaller models for phones and computers. Woof fits the local AI approach. The report says it has 31 million parameters. Parameters are numbers adjusted while a model learns. The 31 million count and the 31.7 MB file size measure different things.
Why it matters
If local speech generation works in everyday use, apps may send less text to cloud servers. That could help people keep more of their information on their own devices. It could also help voice features work when a connection is unavailable. TechCrunch says a model this small is unlikely to replace larger models. It may still support local assistants or personalized podcasts. The important question is not simply whether Woof is small. It is whether the model is small enough, useful enough, and easy enough to run.
What we can confirm
The public demo shows a browser game with a typewriter and a music tuner. The typed text is spoken aloud. The company has stated the model’s size, parameter count, offline claim, and download availability. These points are reported claims, not independent performance results. TechCrunch also notes that Underdog released Bark, a 167 MB speech-to-text model, a few weeks earlier. Bark is meant for transcription and dictation. Underdog has also released a 27-billion-parameter model that runs on laptops. It has a waitlisted app positioned as a local alternative to cloud services.
What remains unknown
The report does not provide an independent test of voice quality. It does not compare Woof with larger models under the same conditions. We also do not know which phones or computers run it smoothly. The article gives no detailed product requirements. It does not explain the terms for commercial use. The offline claim may also vary by device and by the exact app.
What to watch next
The next useful evidence will be outside testing, device requirements, and real apps using Woof. Watch whether local assistants or podcasts become practical. If they do, small models may become useful for different reasons than large models. The demo is a strong signal of direction, but not proof of a finished product.
A tiny voice AI may run on your device
📰 Full story: A 32 MB voice AI brings local speech to the browser
Underdog says its tiny voice model can work close to the user.
Woof
A small AI that reads written words aloud.
local AI
AI that runs on the user’s own device.
parameters
Numbers changed while the AI learns.
💡 The gist
- Woof, a small voice AI, reads typed words aloud.
- It is 31.7 MB, or about 32 MB.
- Underdog says it can work without internet.
Underdog, a startup that builds local AI, made a browser game. The game uses Woof, a text-to-speech model. You type words on a pretend typewriter. Woof says them aloud. A music tuner can make different kinds of music.
Many AI models are large. They need powerful computers to run. Smaller models can fit more easily on phones or laptops. Woof follows this smaller, local approach.
A file size and a parameter count are different. File size describes storage space. Parameters are numbers changed during learning. A smaller file may be easier to carry. It still needs testing for natural speech and speed.
The reason this matters is control. If Woof really works offline, words may stay on your device. They may not need to travel to a cloud server. This could help private notes, local assistants, or personal podcasts. It could also help when internet service is unavailable.
Still, a small file does not prove high quality. The article does not give an independent sound test. It does not show Woof beside a larger model. We do not know which devices run it smoothly. We also do not know the exact product or commercial-use rules.
Underdog has made other models too. Bark is a 167 MB model that turns speech into text. Another model has 27 billion parameters and runs on laptops. These releases show the company’s focus on AI that runs nearby.
The next step is real testing. Watch for comparisons, device requirements, and useful apps. A game demo shows possibility. It does not prove a finished tool.
A little AI can talk
📰 Full story: A 32 MB voice AI brings local speech to the browser
A tiny AI can speak inside a game.
Woof
The name of the AI that reads words aloud.
MB
A way to count computer size.
demo
A small test show.
Underdog (a company that makes AI tools) made Woof. Woof (an AI that reads words aloud) is small. It is 31.7 MB. MB is a way to count computer size. A little game shows Woof working. You type words on a pretend typewriter. Woof says the words aloud. A music knob makes different music. The company says Woof works without internet. That may keep words on your device. But this is only a demo. A demo is a test show. We do not know how well Woof works everywhere. We also do not know if it sounds better than bigger AI.