I've worked on my last gig for a startup that did speech-in-noise recognition and there we've used a recurrent neural net approach to separate background noise from speech. This model was trained on many many hours of audio data as well.