yes, you’re right, the 2nd part maps the features to a final output matrix of shape BxTxC.

Source: Deep Learning on Medium

yes, you’re right, the 2nd part maps the features to a final output matrix of shape BxTxC. To get the probabilities, you have to apply the softmax() function, which is done directly in the CTC operation of TensorFlow, therefore no need to do this manually.

F,B,T,H and C: Forward, Backward, Time (i.e. along width dimension of image), Hidden units (of RNN) and Chars (to be more precise: number of visible chars + 1 special char called “CTC blank”).