A SCORE-TO-SINGING VOICE SYNTHESIS SYSTEM FOR THE GREEK LANGUAGE
Varvara Kyritsi, Anastasia Georgaki, Γεώργιος Κουρουπέτρογλου · University of Michigan Library Repository · 2007
ABSTRACT In this paper, we examine the possibility of generating Greek singing voice with the MBROLA synthesizer, by making use of an already existing diphone database. Our goal is to implement a score-to-singing synthesis system, where the score is written in a score editor and saved in the midi format. However, MBROLA accepts phonetic files as input rather than midi ones and because of this, we construct a midi-file to phonetic-file converter, which applies irrespective of the underlying language. The evaluation of the Greek singing voice synthesis system is achieved through a psycho acoustic experiment. The thirty participants are given two samples and are asked to evaluate them according to vocalness, naturalness, intelligibility and expressivity, in a 1-5 scale MOS test. The results are encouraging, with naturalness and expressivity being ranked above 2.5 in average. Thereafter, we create a new diphone database, from the same text corpus but with a different human voice. The text is recited at constant pitch. After repeating synthesis and evaluation, the quality of the produced voice is evidently improved, with naturalness and expressivity being ranked above 3 in average.