Gradium Open-Sources Phonon, a 31M-Parameter On-Device TTS Model Under Apache 2.0

Gradium, a Paris voice AI company spun out of Kyutai, open-sourced Phonon, a family of small text-to-speech models built to run on the device. The English Phonon 31m ships under Apache 2.0 in a 64.5 MB file and runs on CPUs, phones, a Raspberry Pi 5 and in browsers, with no GPU required.

That changes the math for anyone building voice into an app. No per-character bill, no cloud round trip, and no user audio or text leaving the device.

We like this release, and the reason is practical. On-device speech has long meant robotic voices or slow startup. Gradium reports Phonon 31m produced first audio in 6.9 ms on a MacBook Air M5 CPU and ran nearly 30 times faster than real time. On a Raspberry Pi 5, it starts in 38 ms. Kokoro, a popular open alternative, took 517 ms on the Mac because it doesn't stream. For screen readers, in-car assistants, offline tools and accessibility features, that latency gap is the difference between usable and annoying.

The accuracy story is smart, too. Gradium tested on a hard set full of phone numbers, emails and postal codes, exactly where TTS tends to stumble, and reports 1.70 percent word error rate versus 3.50 percent for Kokoro. Much of that comes from built-in text normalization. On an easier benchmark, Kokoro edges ahead. So this isn't a clean sweep, and to be fair, these are Gradium's own evaluations, including blind listening tests with 1,189 raters.

The licensing deserves a careful read. Only the English 31m model is Apache 2.0. The multilingual 40m models, covering French, German, Spanish and Portuguese, are non-commercial. That's a reasonable business choice, but developers building products outside English will need to talk to Gradium.

The downside is the one every cheap, good voice model carries. Gradium keeps voice cloning and voice design inside its own studio, then lets the voice run locally. That's a sensible guardrail, but natural speech that runs free on any device lowers the cost of robocalls, spam audio and impersonation at scale. As synthetic voices get smaller and better, detection and consent norms need to keep pace.

I'm watching whether small, local models like this become the default for routine speech, leaving cloud APIs for premium voices and cloning.

Phonon uses a compact transformer with a flow-matching head and a reduced version of Kyutai's Mimi decoder, distilled from a larger Phonon teacher model. The runtime covers Rust, Python, Swift, C, C++ and the browser, and the models load in tools including Transformers, MLX-Audio, Ollama and LM Studio.