The majority of the APIs mentioned are probably using DNNs. But those are all online-only, so higher-quality offline engines would definitely be an improvement. I wonder how much effort it would require to integrate them into the SpeechRecognition package.