The Blizzard Challenge 2007: Workshop

in conjunction with the
Sixth ISCA Workshop on Speech Synthesis

Bonn, Germany
August 25, 2007


Non-uniform Unit Selection through Search Strategy for Blizzard Challenge 2007

Feng Ding, Jari Alhonen

Nokia Research Center Beijing, Beijing, China

This present paper describes a non-uniform unit selection synthesis system for the Blizzard Challenge 2007. Non- uniform unit is used to maximize the length of unit sequence to be selected for the target sequence. In a minor modification from the previous implementation, a different search strategy is introduced to transfer a usual phoneme-based speech system to a non-uniform unit selection system, without big changes to the voice database. The front-end analysis results, such as syllable boundary, word boundary, and prosody phrase boundary, are utilized to search from different layers. The probable best small unit instance will be selected, gradually growing up to a longer unit. It is still possible to give up an original phoneme sequence existing in the database if that sequence mismatches the context significantly.

Full Paper

Bibliographic reference.  Ding, Feng / Alhonen, Jari (2007): "Non-uniform unit selection through search strategy for Blizzard Challenge 2007", In BLZ3-2007, paper 012.