Using two-level morphology as a generator-synthesizer interface for concept-to-speech