Multi-tier Non-uniform Unit Selection for Corpus-based Speech Synthesis
Jinhui Yang, Zhiwei Zhao, Yuan Jiang, Guoping Hu, Xiao-Ru Wu · 2006
In this paper, a corpus-based speech synthesis system KB2006 was developed using the speech database provided by Blizzard Challenge 2006. We proposed a novel unit selection method called multi-tier non-uniform unit selection in our corpus-base speech synthesis system. Non-uniform unit (NUU) in our system was defined as a unit sequences that contains one or more joint phoneme units. By using CART algorithm, NUUs with the same phoneme sequence in the inventory were clustered into different classes according to their prosody and acoustic difference. In the unit selection stage, a multi-tier NUUs selection algorithm was adopted by treating different NUUs with several criterions. With the discrimination, proper candidate units that close to the target unit can be selected for speech concatenation. 1.