On the pronunciation of unknown words by analogy in text-to-speech systems: an evaluation