WaveNet: A Generative Model for Raw Audio - https://arxiv.org/abs/1609.03499:
This is a very very very interesting paper. Like AlphaGo, it was very pleased to read the paper. Abeit I also can imagine the framework is super expensive to train!
The techniques themselves are not new. Yet to make all things work, it would require significant expertise that Deep Mind researchers are second to none.
What a time to be alive! Deep Learning is changing lots of things, turning impossible to possible.
Ps: The experiment about speech generation is very cool. It reminds me of this http://karpathy.github.io/2015/05/21/rnn-effectiveness/