Prosody Correction Preserving Speaker Individuality in English-Read-By-Japanese Speech Synthesis Based on HMM

To build an English acoustic model that well captures speaker individuality of each Japanese speaker, a framework using English-Read-by-Japanese (ERJ) voices is effective as it enables to directly model speaker-dependent acoustic characteristics. However, naturalness of English speech synthesized by such an ERJ acoustic model is significantly degraded as it is directly affected by prosodic differences and pronunciation errors often caused by differences of a language system between Japanese and English. To synthesize more natural English speech while preserving speaker individuality of individual Japanese speakers, we propose a technique to correct prosody of ERJ voices based on that of a native English speaker. The duration and power of the native English speaker are effectively used to develop the ERJ acoustic model for each Japanese speaker by using model adaptation techniques in HMM-based speech synthesis. The experimental results show that our proposed method is capable of significantly improving naturalness of ERJ synthetic speech while preserving its speaker individuality.