On-device text-to-speech with all 7 Phonon models (31m English + 40m en/fr/de/es/pt/multilingual), ~100 voices, CPU-only, no PyTorch