repeated sequence of characters

[MrTTTTTwinkel]

When I use the trained model to predict a text image (the dataset I used was 100W generated images + 3W labelled real images), say the real result should be: happy, however the predicted result is: hhhaappyy. Any suggestions on the cause of this phenomenon and what I should do to solve this problem? Thanks a million :)