Gradium's Phonon: 65 MB text-to-speech that runs on-device

Gradium, a speech-AI company, says its Phonon text-to-speech models run entirely on the user's device, with no server and no network. The company set out the details in a Hugging Face blog post published on 5 November 2026, two weeks after opening the weights. The post is vendor-published, co-written with Hugging Face staff, and the figures below are Gradium's.

What was released

• Two models of about 65 MB and 75 MB. Phonon 31m speaks English; Phonon 40m covers English, French, German, Spanish and Portuguese.
• More than 100 built-in voices, plus custom voices through voice cloning and voice design in Gradium Studio.
• Gradium says the models handle phone numbers, dates and codes, which it describes as a common failure point for other text-to-speech systems.
• A new inference framework, ptts, with more than 10 backends across CPU (Intel, Arm, WebAssembly), GPU (CUDA, Vulkan, Metal, WebGPU) and NPU (Apple Core ML, Snapdragon). The model is also in Hugging Face Transformers from v5.21, aimed at batched GPU generation such as building speech datasets.

The licence is the catch

Phonon 31m is Apache 2.0, so a company can ship it in a commercial product. The other models are CC BY-NC, free only for non-commercial projects. For commercial use of those, Gradium says pricing is based on the number of users rather than the amount of speech generated. The post gives no price. If you need French, German, Spanish or Portuguese in a paid product, that is a conversation with Gradium.

What the numbers say, and what they leave out

Gradium says Phonon was downloaded more than 200,000 times in its first two weeks. On Hugging Face's Open TTS Leaderboard, which covers 23 languages and scores intelligibility by word error rate, speed, and naturalness, the post says Phonon ranks first in German and Spanish on intelligibility and in the top three in every language Gradium evaluated. It also says the models lead the CPU streaming benchmark on time-to-first-audio.

The post does not give the word error rates themselves, the hardware used for the speed test, or any comparison on a business task such as call-centre prompts or product names. It also says training was initially numerically unstable and needed several tricks to work across all five languages, a reminder that voice quality in a language you rely on is worth checking yourself. Naturalness is scored with automated measures (UTMOS for English), not listeners, though the leaderboard has a Listen tab for comparing outputs by ear.

Why a team might care

Running speech on the device means customer text never leaves it and there is no per-character cloud bill. Gradium's examples include a coffee-brewing guide on an Apple Watch with Wi-Fi off and a rep counter on an iPhone with no signal. Those are demos, not deployments, and the post names no company running Phonon in production.