The ModelTalker System

The ModelTalker TTS system has recently been largely rewritten to change its design to make full use of information derived from talker-specific HMMs that are trained in optimizing phonetic transcription and alignment. Because this system is substantially different than the system last used in a Blizzard challenge in 2005, we decided to participate once again. This allows us to both compare performance with the previous version on similar tasks, and with the latest cutting edge TTS technology. The current version of ModelTalker appears to be comparable with the previous version in segmental intelligibility and substantially improved in the naturalness of its synthetic output.