Gradium open-sourced its on-device TTS

Gradium released Phonon, an open-weight on-device text-to-speech suite built to run offline on phones, laptops, single-board computers and in browsers. It ships as three compact models - Phonon 31m (31M parameters, 64 MB, Apache 2.0) and two 40M variants (74.6 MB each: single-language and multilingual, CC BY-NC) - that target naturalness and accuracy with low word error rates (Phonon 31m reported 1.7% WER on difficult sentences). Phonon supports English, French, German, Spanish and Portuguese, produces WAV output locally, and claims high efficiency (up to ~22× real-time on recent Apple silicon). The emphasis is privacy-first: text and audio never leave the device once models and voices are downloaded.

Deployment and developer tooling are broad: a Rust CLI plus examples in Python, Swift and browser builds; runtimes include x86-64, ARM, Apple Neural Engine/Core ML, WebAssembly/WebGPU, CUDA/Metal/Vulkan and even Raspberry Pi 5. Phonon is demonstrated in demos and apps (voice agents, game voices, watch, Chrome extension) and integrates with Gradium Studio/API for voice cloning or voice design; downloaded voices run locally like any other voice. The code and models live on GitHub (gradium-ai/xn-ptts), commercial use of CC BY-NC models requires contacting sales, and several industry partners are listed as early adopters.