From the source
Gradium released Phonon, an open-source on-device text-to-speech model that generates speech locally on phones, laptops, embedded hardware and in the browser, with no network round-trip or cloud dependency.
- The company says Phonon has 31m parameters against 82m for Kokoro, the most widely used open on-device TTS model, and that it wins on overall preference in blind human evaluations ahead of Kokoro, Magpie and Pocket TTS.
- Two variants ship at launch, one with 40 million parameters and one with 31 million, sized to fit inside a mobile app or load in a browser tab.
- A single multilingual model covers English, French, German, Spanish and Portuguese, with more languages on the roadmap, and it comes with a custom inference codebase that Gradium says is faster than Kokoro and llama.cpp-based runtimes on the same hardware.
- On desktops and servers with
NVIDIA GPUs, the CUDA backend runs inference 8x faster than on CPU, generating the first bit of speech in 5ms on an NVIDIA L40S against 43ms on an EPYC 7413.
