This paper presents a new approach to synthesizing fast

speech in unit selection synthesis. After recording two inventories - one at normal and one at fast speech rate articulated as accurately as possible - speech was synthesized from both corpora independently. Since fast speech differs from normal rate speech in terms of acoustic characteristics, the concept of multi-phone (phoxsy) units proposed by Breuer and Abresch [1] was explored for synthetic speech generated from both

speaking rate inventories. A perceptual evaluation showed that phoxsy units enhanced the intelligibility especially for fast synthetic speech significantly.