Line-level text recognition for historical print and handwriting.
Note Newest: subword decoder, read with use_decoder=True.
Note CTC and KenLM only, smaller and faster.