I’m not sure exactly, but it’s possible you can achieve what is documented in the paper by adapting the KenLMBeamScorer code, but I don’t remember the whole context of the paper, and it’s DeepSpeech 2 and not DeepSpeech, so maybe it’s not a valid suggestion from the authors in our context.
But definitively, if you need to play around the CTC decoding step, then this is likely the best way to do it.