a question about CTC implemention

[xray1111]

I noticed that the forward-backword algorithm in ctc implemention has a little difference from Alex's paper: it only consider the previous nearest 2 neighbor nodes when computing the forward variable, while in Alex's paper the including number of neighbor nodes could be 2 or 3, which depends on current node's character (formula 7.16 "Supervised Sequence Labelling with Recurrent Neural Networks").

Is this because OCR's input sequence is much longer than output label, and we couldn't guess what the output character is when 1 pix column of the image matrix is feed in?

Thanks very much!