Removes unused keyword-spotting model. Makes on-device voice model a runtime download rather than bundling it. Removes native-code library versions for 32-bit and x86_64 processors.