Your code looks heavily to rely on https://github.com/keras-team/keras/blob/master/examples/image_ocr.py but you do not seem to be the original author (and me neither).
Please quote the original authors.

Hi, i've two issues with the code i hope you might help me with.

First: Training for the 25 (20+5) epochs does not result in accurate predictions, i ran this both locally and on the google collab notebook, not changing anything, and both times the predictions were really bad.
basically the output was maybe a few letter from the text repeated up to the max length.
did you change anything from the training to the notebook? or did you train longer?

Second: (this might be a bit more on my side, but you might have a clue)
Loading the provided weights on the google collab notebook results in accurate predictions, however loading them locally drops the accuracy. Meaning that it gets most words right, but with a letter being wrong, this applies to beam search as well, and the probabilities for the letters are a bit more "diffuse", not as sharp as the ones on the google collab version.
So basically i'm running the same code, loading the same weights, but having a decrease of accuracy locally. Understandable if you can't help me there, just thought you might have run into something like this before.

Locally I'm on a rather fresh install of Ubuntu, all prerequisites installed, using an anaconda virtualenv.

Thanks for the help,
Nic

ps: i also tried a fresh anaconda env in case something was interfering, with no change