Vocaloid V4
I experimented with vocal recordings and UTAU to come up with some suggestions for V4:
• Assimilation and co-articulation should improve, particularly for vowels.
•Stress needs to become easier to manipulate and control.
•Sounds that are cut short need to sound smoother. One hypothesis for achieving this is to create two voice samples for each sound--one at a normal length and one shorter. This could get tricky, however.
•Most importantly, voice banks need to be more extensive. Triphones obviously improve the quality of pronunciation, so V4 banks should have more of them. Ideally, V4 should experiment with the inclusion of whole morphemes and words in voice banks, moving slowly into a system of unit selection synthesis rather than diphone synthesis.
Diphone synthesis is icky. Don't the makers of Vocaloid have computational linguists on their team to tell them that?
