Gradium, a Paris-based startup spun out of the Kyutai research lab, has released Phonon, a family of text-to-speech models designed to run directly on the device. The announced versions have 31 and 40 million parameters, both under 50 million, and are released as open source together with the inference code.
For developers, the key point is that the model needs no connection: text and audio stay on the phone, the laptop or the browser. According to the model card on Hugging Face, the English version generates 24 kHz speech and streams it while still producing it, at a size that fits in a browser tab, on a smartphone or on a laptop CPU.
Image: phonon.studio
Phonon builds on Pocket TTS and the Mimi codec, both developed by Kyutai. The inference side is open too: the xn-ptts project on GitHub describes it as built for offline assistants, accessibility tools and interactive applications that need speech without a cloud round trip. The stack targets more than ten backends, including CPUs, GPUs and NPUs, and is mostly written in Rust.
The most concrete figure for anyone evaluating adoption is portability. A model under 50 million parameters fits on devices that cannot run large cloud voice systems, and it reduces dependence on external APIs for voices used in apps, home devices or accessibility tools. In its presentation, Gradium speaks of a new standard for naturalness and accuracy for local execution, with broad language support; the sources provide no independent measurements confirming the comparison.
The open-source release fits a trend already visible in our archive: in a few months local voice models such as Qwen3-TTS in llama.cpp and Chatterbox-Nano in the browser have arrived. Phonon, however, goes much lower in size, which makes it a natural fit for devices with little memory.
Gradium had already opened a commercial line with its Voice Design tool, which generates voices from text descriptions. Phonon is the company's first open-source step, and the company says it plans further releases.
No comments yet. Be the first!