Introducing Leokani-tts
Frontier-quality speech on any CPU.
AI measured in milliwatts, not megawatts.
Most of today's AI runs in data centers: rooms of GPUs, cooling and power lines measured in megawatts. Every question you ask and every sentence you hear travels there and back. We think a lot of everyday AI doesn't need to work that way. It can run on the devices you already own: your phone, your laptop, a single-board computer, even a microcontroller. That means no data center, no GPU, nothing leaving your device, and far less power.
That's the whole idea behind Milliwatt Labs. Leokani-tts is the first thing we're shipping on that road.
What it is
Leokani-tts turns English text into natural speech. It has 7 million parameters, and the voice file is 14 MB. The name is Hawaiian: leo means voice and kani means sound.
In blind listening tests against the much larger model it learned from, it was rated the same on almost every line.
What it runs on, measured
| Where | Speed | First sound |
|---|---|---|
| Mac (M3 Pro, native app) | 296× faster than real time | 19 ms |
| iPhone 17 Pro Max | 101× | 64 ms |
| A $50 Android phone (moto g play 2024) | 20× | 234 ms |
| Raspberry Pi (CM4) | 11× | |
| A 1-vCPU Linux server | 36× | |
| Your web browser, no server | 97× | 64 ms |
"Faster than real time" means a 10-second sentence takes a fraction of a second to make. Every one of these runs on the CPU or the phone's own chip. None of them needs a GPU.
You can try it right in your browser, download it for Mac, Linux, Raspberry Pi and Android, or drop it into your own app as a JavaScript library.
What's next
- Smaller. An 8-bit version already sounds the same in blind tests at half the size, and it's headed for a 32 MB ESP32 microcontroller next.
- More expressive. We're teaching Leokani to dial emotion up and down, from calm assistant to lively storyteller, without losing its voice.
- More than voice. Speech is one piece. We want the whole loop of listening, thinking and answering to run on the device in your hand.
It's early, and Leokani isn't perfect. But every model we ship is meant to prove the same point: useful AI doesn't have to cost megawatts.
A sustainable AI future.
Leokani-tts is free for non-commercial use (CC BY-NC 4.0). For commercial use or a custom voice, contact hello@milliwattlabs.com.