The IBM Submission to the 2006 Blizzard Text-to-Speech Challenge

In this paper, we present two concatenative text-to-speech systems built from the “Blizzard Challenge” speech databases. The two systems differ primarily in their segment selection cost function. One system has our baseline cost function, and the other has a cost function which has been altered to potentially better handle small datasets. Results indicate that both systems perform similarly in terms of MOS and intelligibility.