HMM-Based Speech Synthesis System with Prosody Modification Based on Speech Input
友理 西垣, Shinnosuke Takamichi, Toda Tomoki, Graham Neubig, Sakriani Sakti, Satoshi Nakamura · IEICE technical report. Speech · 2014
As a creative activity using speech synthesis technologies has been grown rapidly, it is desired to develop an interface to synthesize speech of a specific target speaker as users want. In this report, we propose a prosody modification method using user’s speech inputs in HMM-based speech synthesis system in order to achieve high-quality and expressive speech synthesis. The propose method allows users to guide prosody of synthetic speech of the target speaker by using their own voices while preserving original functions of the HMM-based speech synthesis system as a text-to-speech synthesis system. Both duration information of the input speech extracted by performing HMM state alignment and F0 patterns of the input speech are effectively used to control the duration and F0 patterns of synthetic speech of the target speaker. To alleviate the degradation of naturalness caused by prosodic mismatches between the input speech and the synthetic speech, we investigate an appropriate unit for the HMM state alignment, model adaptation for building an HMM used for the alignment, and correction of unvoiced/voiced information. Experimental evaluations are conducted for multiple input speakers, which demonstrates the effectiveness of the proposed method.