Speaker-TTS voice mapping towards natural and characteristic robot storytelling

摘要

Robot storytelling has the potential for its practical use in various domains such as entertainment, education, and rehabilitation. However, relying on human-recorded voices for natural storytelling is costly, and automation with text-to-speech systems is not readily applicable due to the difficulty of reflecting the full nature of stories in TTS systems. In this paper, we address the problem of automating robot storytelling with a particular focus on two issues: speaker identification and speaker-TTS voice mapping. We first conduct text analysis with rich linguistic clues to identify speakers from a given textual story. We then consider the task of speaker-TTS voice mapping as the graph coloring problem and propose effective algorithms for assigning voices to speakers given a limited number of TTS voices. Finally, we perform a user experiment on validating the usefulness of our method. The results demonstrate that our system significantly outperforms baseline systems and is also more acceptable to users.