CNN+LSTM OCR model not predicting “is” correctly

I have made a model in tensorflow with the following structure for OCR. I trained it on a significantly large dataset(250K images) and also getting around 94% accuracy. But the model predicts "is" as "1s" or "1S" most of the times. I checked the dataset and it contains quite a lot of instances of "is" so I dont think dataset is the problem. Also the model predicts correctly when it is a part of another word such as "this" or "christian". What could possibly be the reasons for this behaviour how can I improve the model to detect "is". I have made a model in tensorflow with the foll