Explicit Prosody Control to Realize Discourse Focus in End-to-End Text-to-Speech

In this paper, we propose an end-to-end text-to-speech (E2E-TTS) system that enables emphasis prosody using explicit symbols. The proposal hinges on the novel idea that emphasis tags can be generated for TTS training by comparing speech readings of novels by a professional voice actor with synthesized speech generated from the same texts. While human speech naturally contains appropriate emphasis for novels, synthesized speech tends to be more monotonous because of E2E-TTS are typically trained using news article speech. Using this comparative method, we can determine the appropriate emphasis prosody. An explicit prosody-controllable TTS (EPC-TTS) is trained with texts and attached emphasis tags. As prosodic features, F0, power, and a combination of the two are examined. Judging from objective and subjective evaluations, we confirmed that the EPC-TTS can successfully realize a phrase focus using explicit symbols, and the combination of F0 and power yielded the best performance.