Image: Gradium

French voice AI company Gradium has released Phonon, a set of small text-to-speech models that run on a phone, a laptop or even in a browser tab with no internet connection. The company announced the release on X and its blog. The smallest model, Phonon 31m, is free to use in commercial products.

A new frontier for on-device speech.

We're open-sourcing Phonon, a suite of TTS models under 50m parameters, setting a new standard for naturalness and accuracy across languages, with the broadest inference support to date.

Phonon-31m is available under Apache 2.0. pic.x.com

— Gradium (@GradiumAI) October 20, 2026

Text-to-speech is the voice that reads out your sat nav directions, your screen reader or an AI assistant's replies. The best voices today usually live in the cloud. Services like ElevenLabs send your text to a server and stream audio back. That costs money per use, adds a delay and means your words leave your device. The voices built into phones run offline but often sound flat.

Phonon tries to close that gap. The English Phonon 31m model has 31 million parameters, a rough measure of a model's size, and its download is about 64 MB. That is smaller than many phone games. Gradium says it starts speaking in about 7 milliseconds on a laptop processor, and it streams audio while it is still working on the rest of the sentence.

Gradium also claims Phonon beats better-known open models. In its own tests on a MacBook Air, Kokoro, a popular free voice model, took 517 milliseconds before any sound came out, according to Gradium's numbers as reported by AICoder. Gradium says Phonon also made fewer reading mistakes on a set of hard sentences full of numbers and dates. Gradium built that test set itself. No independent lab has checked these results yet, so treat them as the company's claims.

The release has three parts, and the licences differ. Phonon 31m speaks English with 37 built-in voices under Apache 2.0, which allows paid products. Phonon 40m comes in separate English, French, German, Spanish and Portuguese versions. A multilingual 40m model speaks all five. Both 40m options use a non-commercial licence, so companies must talk to Gradium to ship them.

The models are not built from nothing. The Hugging Face model page says Phonon builds on Pocket TTS and the Mimi audio codec from Kyutai, a Paris research lab. Gradium says it optimised the models for more than 10 hardware types, from Intel and Arm chips to Apple and Snapdragon phone processors.

This matters most to people who build apps. A developer can now add a natural voice to a reading app, a game or a smart home speaker without paying per word or sending user text to a server. For ordinary users, that could mean screen readers and offline assistants that sound more human and keep what you type on your device. It could also help people in places with weak or costly internet.

There are limits. The model card says Phonon speaks only its listed languages, and long texts are read sentence by sentence, so the tone can drift. Gradium also asks users not to clone a voice without the speaker's consent or pass synthetic speech off as a real person.

How to try Phonon

Developers can download the weights from Hugging Face and the code from Gradium's GitHub. Anyone with Python's uv tool can make a test clip with one command, uvx ptts. The repo also includes a server that works with apps built for OpenAI's speech API. Pick Phonon 31m if you plan to sell what you build.