Comment by trowngon

Irrespective of the subject Deepspeech is very old archtecture with suboptimal results. You'd better try any recent conformer implementations (flashlight, nemo, wenet, etc) or wav2vec.