Vowel Synthesis Based on the Spectral Morphing and Its Application to Speaker Conversion

In this paper, we propose a speaker conversion and morphing technique for speech synthesis based on the spectral morphing, although the target synthesized voice is restricted to vowel and vowel-like speech sound at present. In the experiments, we have prepared the databases of four speakers, which consisted of the normalized spectra on Japanese fundamental five vowels and the mean fundamental frequency (pitch). The raw speech parameters including three formant frequencies, pitch pattern and amplitude one were extracted from the connected vowels, /aiueo/ uttered by a male and converted or morphed into another speaker's voice using each database. As results of the dissimilarity tests, the desirable performances in speaker conversion or speaker morphing were confirmed in the proposed vowel synthesis. It is expected that this type of voice synthesis method will be applicable to a hybrid TTS, where a specific speaker's voice is synthesized according to the synthetic rule