From the source
Gradium is open-sourcing Phonon, a family of 31M and 40M parameter text-to-speech models built to run on-device, from a private beta shaped by early users and partners.
- Until now, on-device voice meant compromise, with limited language coverage, models too large for the memory budget, or quality too low to ship, the company says.
- The English Phonon 31m starts speaking in 7 ms on a laptop CPU and ships under Apache 2.0, and Phonon 40m comes as language-specific variants or one multilingual model with a single set of weights.
- The multilingual model covers English, French, German, Spanish and Portuguese, and all models run on CPU, GPU and NPU across Intel, Arm, WebAssembly, CUDA, Vulkan, Metal, WebGPU,
Apple Core ML and Snapdragon.
- Phonon 31m occupies 64 MB and Phonon 40m 74.6 MB; Gradium also offers custom models for commercial use.
