In my experience DeepSpeech with the default Mozilla model + 5-gram is extremely sensitive to background noise. Fiddling with the hyperparameters, e.g. upweighting the language model, helps a little.