Speech Synthesis Based on Hidden Markov Models