Gradium open-sources Phonon, a 31M-parameter on-device TTS model

Paris-based Gradium, spun out of Kyutai, has released Phonon, a family of small text-to-speech models that run locally. The English Phonon 31m ships under Apache 2.0 in a 64.5 MB q8 GGUF file and runs on CPUs, phones, Raspberry Pi 5 and browsers without a GPU.

• Phonon 31m: 31M parameters, 64.5 MB, 37 voices, Apache 2.0
• Phonon 40m: 40M parameters, 74.6 MB, CC BY-NC 4.0 license
• About 7 ms to first audio on CPU, 3.7x real time on Raspberry Pi 5
• 1.70% WER on a hard set versus 3.50% for Kokoro