Tiny Free AI Makes Any Gadget Talk — Even a $3 Microchip

Imagine a doorbell that doesn't just buzz but says “Delivery at the front door” in a natural voice — and costs almost nothing to make. A developer just released sanoTTS, a text-to-speech system (AI that reads text out loud) so compact it can run on a $3 computer chip with less memory than a 1990s video game console. The smallest version takes up just 337KB when compressed, small enough to fit inside cheap electronics. And it's open source, so anyone can use it.

Why does size matter so much? Today's talking assistants like smart speakers usually depend on huge data centers in the cloud. That means your words travel over the internet, wait for a giant AI to answer, and burn energy along the way. sanoTTS changes the math. The system has 11 voices, supports 6 languages, and on that $3 chip it produces 4 seconds of speech in 1 second. It is about 1,000 times smaller than some popular voice models, yet in quality tests its larger version scores better than speech AIs that are 3 to 10 times heavier. In other words: nearly free hardware can now sound better than software that needs a miniature server room.

What can you do with this? Think of devices that never needed a screen or speaker before — a pillbox that reminds you to take your medicine aloud, a thermostat that tells you the daily forecast, a vending machine that offers help in multiple languages. Because everything stays on the device, your commands aren't sent to a company's servers, which also helps with privacy. sanoTTS also works in a web browser, so you can try it without installing anything.

Of course, there are limits. A few hundred kilobytes of software can't match the expressive emotion, accents, or crispness of billion-dollar cloud voices. The current voices are clear and pleasant, but they aren't audiobook narrators. Still, as the developer notes, this is the smallest complete neural text-to-speech stack of its kind. The next time a cheap gadget talks to you, this is likely why.