Kale TTS tiny: 45.7M params, 160ms first sound, frozen Ettin sentence reader
NERDDISCO · x · 2026-11-20
Kale TTS tiny is the clearer, more expressive variant: 24x realtime, 160ms to first sound on device, 45.7M parameters, 79.9MB on CPU (183.2MB WebGPU). Unlike nano it adds a frozen Ettin word reader (16.8M, from Orion Weller and JHU CLSP) so it understands whole sentences; the 24.8M planner is reused frozen from kale-tts-1-base and the 4.1M sound generator learns from Maya-1 audio.