Do text-to-speech synthesisers pronounce correctly? A preliminary study

This paper evaluates 4 commercial text-to-speech synthesisers used by dyslexic people to listen to and proof read text. Two evaluators listened to 704 common English words and determined whether the words were correctly pronounced or not. Where the evaluators agree on incorrect pronunciation, the proportion of correct pronunciations for the four synthesisers is in the range 98.9% to 99.6% of the 704 words. The evaluators also listened to the same synthesisers speaking phrases in which there were 44 pairs of homographs and determined whether each instance of the homograph was correctly spoken or not. The level of correctness for the four synthesisers ranged from 76.3% to 91.3%.