I believe they are focused on the hybrid approach which replaces the corresponding component with DL.

For Listen, Attend and spell, some people use Kaldi for feature extraction. You can go to other open sources to implement the rest.