How to build a character based seq2seq tensorflow model for spell correction?

[Murtuza Dahodwala]

I am trying to build a character based seq2seq model using tensorflow for food item names correction. The input/output would be like

IN: Cheessseee Quisadillas

OUT: Cheese Quesadillas

I have tried training a word level seq2seq but it does not give much good results. I have also read that we only used attention mechanism on long sequences and I have a Maximum of 6 word sentence(Item Name).

Apart from this, Can I use a language model like BERT or GPT for such task

Can anyone please suggest the changes or provide better resources for implementing the same.?