DNNs have lead to massive improvements in speech, specifically in the acoustic model. RNN based language models interpolated with modified kn smoothed ngrams (often plus some other cool stuff) are also the state of the art, although they're often too slow to use in practice.