Synthesizing variation in prosody for Text-to-Speech