The Lessac Technologies System for Blizzard Challenge 2010
Rattima Nitisaroj, Reiner F. Wilhelms-Tricarico, Brian Mottershead, John Reichenbach, Gary R. Marple · 2010
For Blizzard Challenge 2010, Lessac Technologies built its first British English voice from the provided full database. To enhance methods for target cost calculation and unit selection, instead of traditional phonetic symbols, we used a more finegrained set of Lessemes to label units and applied the Hierarchical Mixture of Experts model to map linguistic features to acoustic parameters. The evaluation results show that we performed relatively well on similarity to the original speaker, and comparable to most systems with respect to naturalness. The high word error rate suggests that we need to improve on signal processing for concatenation. Index Terms : speech synthesis, Blizzard Challenge, Lesseme