Tiny neural text-to-speech (294k-2.27M params), 9 voices in English, Indonesian and Vietnamese, CPU-only