Opera Singing Voice Synthesis Considering Vowel Variations

In order to improve a synthesized opera singing voice that is obtained using deep neural networks (DNN), we focus on the vowel variations of the opera singing voice. Variations in the F0 of the vowels of the opera singing voice are analyzed using k-means clustering. Then, from the analysis results, a new score feature for a DNN-based voice synthesis system is introduced as a one-hot vector. The objective evaluation results show that the use of the new score feature improves the quality of the synthesized opera singing voice.