Prosody synthesis by unit selection and transplantation on diphones

Corpus-based selection and concatenation synthesis can be used for prosodic generation in a small size diphone synthesis system. It is shown that we can do without syntactic analysis and prosodic rules: only the prosodic part of the corpus is retained, and not the speech signal itself. A comparative evaluation test is then conducted, using a same MBROLA diphone system for segmental synthesis. It shows that the prosody obtained by selection rated better than the prosody computed by our previous prosodic rules, but worse than a better prosody by rules and natural speech. As a matter of fact, the best prosody by rules outperforms natural speech on this task.