Phonon TTS tops a new speech leaderboard. The better model isn't the free one
Gradium's Phonon TTS posts the fastest time-to-first-audio on Hugging Face's new Open TTS Leaderboard. The catch is which of its two models you're actually allowed to ship.
Hugging Face published a retrospective on November 5 about a model that actually launched two weeks earlier. The post is written by Gradium’s own team, same as always, but this time the numbers come from a leaderboard Hugging Face built, not just Gradium’s own test set. Phonon is Gradium’s on-device text-to-speech model.
Phonon runs on the device making the request. Not a server. No network round trip, no audio leaving the phone or laptop it’s generated on. That matters for screen readers, offline assistants and anything that can’t wait on a cloud call, and it’s been the pitch since Gradium, the voice AI company founded by former Kyutai researchers, first announced the project in April.
What actually shipped
Phonon comes in two sizes, and they’re not licensed the same way.
| Model | Size (q8) | Languages | License |
|---|---|---|---|
| Phonon 31m | ~65 MB | English | Apache 2.0, free for commercial use |
| Phonon 40m | ~75 MB | English, French, German, Spanish, Portuguese | CC BY-NC, non-commercial only |
Gradium’s own release post names the 40m model as the one to pick when expressivity and accuracy on hard text matter most. My read: that’s a polite way of saying it’s the better-sounding one.
Two models, one catch
So the model Gradium calls the stronger of the two isn’t the one you can build a product on without a licensing call to their sales team. A developer who wrote about the release called this a “licence trap”: the launch framing is “open-sourcing Phonon,” and that’s technically true of the 31m model, but the family framing blurs which model that openness actually covers.
What the leaderboard adds
The November 5 post is Gradium reporting its own results on Hugging Face’s Open TTS Leaderboard, a project the company didn’t build. Across 23 languages, the leaderboard scores models on intelligibility (word error rate), speed and naturalness. Phonon ranks first on intelligibility in German and Spanish only. Top three everywhere else Gradium tested, against models it says run up to 100 times larger.
Speed is where the numbers are starkest. On the leaderboard’s CPU streaming benchmark, Phonon 31m starts speaking in 28 milliseconds and the 40m model in 38 milliseconds, against 165 milliseconds for Pocket TTS, the model Phonon’s own architecture builds on, at the 95th percentile. Fast. Gradium’s own release post put the 31m model’s time-to-first-audio even lower, at 7ms on a MacBook Air M5 CPU, against 517ms for Kokoro, the open on-device TTS model Gradium says is currently the most widely used. The two numbers aren’t measuring the same hardware, so don’t stack them, but they point the same direction.
Where I land
My read: Phonon landing in the top three on intelligibility against models Gradium says run up to 100 times larger, on a leaderboard it didn’t build, is a real result, not just a marketing number, because Hugging Face’s leaderboard is neutral ground. The accuracy claims hold up less cleanly. Gradium’s word-error-rate numbers come from its own 500-sentence test set, built from cases developers sent the company, which it has also published, so it can be checked, but it’s still Gradium’s own set, not an independent one. Judge the latency numbers more than the preference numbers.
My read on where this goes: if Phonon’s adoption keeps up, not just its benchmarks, on-device speech stops being a fallback for when the cloud option is too slow or too expensive, and starts being the default for anything that runs offline. Gradium says Phonon was downloaded more than 200,000 times in two weeks, and Hugging Face’s own Transformers library added native support from version 5.21. My take on the open part of “open-sourcing Phonon”: whether most of those builders end up stuck on the free, English-only model or pay Gradium for the better one is the part the launch posts don’t dwell on.
Related: Reflection AI’s open-weight language model Beam, where the same gap between a generous headline and a narrower reality showed up in a different open-weight release. A different company hit the same free-tier-vs-better-model split a day later with Underdog’s 32MB Woof text-to-speech model.
Sources
Note on sources: I read Hugging Face’s and Gradium’s own posts in full, plus the dev.to write-up that flagged the licensing split. All four are linked below. The Hugging Face post is Gradium-authored, not an independent review, so its benchmark framing is still the company’s own. I didn’t independently re-run the live leaderboard or Gradium’s test set myself, and Gradium’s two October posts about the launch put the date a day apart from each other.
- A Closer Look at Phonon TTS (Hugging Face / Gradium, November 5, 2026)
- Open-Sourcing Phonon: A Suite of Frontier On-Device TTS Models (Gradium, October 2026)
- Phonon TTS: 64 MB, Six Milliseconds, One Licence Trap (DEV Community)
- Gradium releases Phonon, an open-source text-to-speech model that runs on any device (Gradium, October 21, 2026)