ISCA Archive Blizzard 2009
ISCA Archive Blizzard 2009

CircumReality text-to-speech, a talking speech recognizer

Mike Rozak

The CircumReality text-to-speech engine's mean opinion (MOS) and similarity-to-original scores have improved significantly over the last three Blizzard Challenges. MOS has increased from 1.3 in 2007 to 2.8 in 2009. This paper describes the algorithmic improvements made to the CircumReality engine between the 2008 and 2009 Blizzard Challenges. The most significant improvements stemmed from a shift in the underlying philosophy of the engine: integrating automatic speech recognition (ASR) into the speech synthesis engine and creating a "talking speech recognizer".