← AI studies
Study · 7 Nov 2026 · 1 min read
A 40 MB voice model beat a 41.5 GB one. Read the fine print.
Underdog Woof is a tiny open-weight text-to-speech model that runs on a laptop CPU. The numbers are striking — and there are three things to check before you swap your TTS stack.
A 40 MB voice model just beat a 41.5 GB one. That's roughly 1,000× smaller, and it runs on a laptop CPU.
It's called Underdog Woof, a new open-weight text-to-speech model from Underdog AI.
The numbers (from their own benchmark)
- 1.42% word error rate on 500 hard English sentences
- Fish Audio Drama 3 (41.5 GB): 1.89%
- ElevenLabs Flash v2.5: 6.63%
- ~14× faster than real time on a single CPU thread
- Needs only NumPy — no GPU, no server, no API key
There's even a 31.7 MB version that runs in the browser.
Before you replace your TTS stack, three things
- Accuracy isn't naturalness. Word error rate measures whether speech is understood correctly, not whether it sounds human. Naturalness is usually where big models win.
- It's a self-run benchmark. Underdog ran and scored every model themselves. There are no independent results yet.
- The licence is non-commercial. Free to download is not the same as free to ship in a product.
Why it still matters
The direction is what counts: voice is moving off the cloud and onto the device. Faster responses, no per-request API bills, and user audio that never leaves the phone. For anyone building voice features, that changes the maths.