Eigenvoice-Based Character Conversion for Arbitrary Speakers Using Various Character Voices of a Skilled Voice Actor

抄録

We describe a new method of voice conversion aimed at character conversion by the eigenvoice Gaussian mixture model (EV-GMM) approach. Using an eigenvoice space built from 273 speakers and speech samples of three different characters created by a single skilled voice actor/actress, the conversion can generate the voices of the three characters from an arbitrary speaker, while retaining the speaker identity. Listening tests were carried out by presenting two kinds of synthetic voices: those before and after the character conversion. The results showed that listeners, both native and non-native speakers, can perceive well the character voice difference intended by experimenters. Moreover, our proposed method performs better than the F₀-based approach. The understanding of how to create synthetic speeches that realize character conversion within the same individual reveals a necessity for and chances to develop a more intelligent correct feedback for voice training as well as a voice therapy system, which will provide participants with an idea of what their expected voices should be.