Gradium released Phonon, a family of open-weight text-to-speech models. They run fully on-device. Three variants exist: 31m, 40m, and 40m multilingual. The smallest model is 64.5 MB. They support English, French, German, Spanish, and Portuguese. They work offline. They run faster than real time on a Raspberry Pi 5. They also run on iPhones, Macs, Android, and in browsers. The 31m model has a 1.7% word error rate on English. Phonon 40m multilingual handles all five languages in one model. They are compatible with Apple, NVIDIA, Vulkan, and WebGPU. They are available under Apache 2.0 or CC BY-NC licenses. Commercial use requires contacting Gradium.
What changed
Text-to-speech models typically required cloud processing or large on-device models.
What it unlocks
Running natural-sounding text-to-speech entirely offline on phones, laptops, and embedded hardware.
- 64.5 MB smallest model
- 1.7% word error rate on English
- 5 languages supported
- up to 22× faster than real time on iPhone 16 Pro
Sources