“Learning distributions from samples” at the Information Theory, Learning and Big Data workshop hosted at the Simons Institute (March 2015). For a five-minute version of the talk with more results, please look at my talk from COLT 2015.
“Improving the cutset bound” at the Between Shannon and Hamming: Network Information Theory and Combinatorics workshop hosted at the Banff International Research Station (March 2015).
Some great advice here. A couple of other concrete ideas I've found useful:
Prefer speaking loudly away from the mic rather than softly in to the mic.
Put the effort to structure your talk into a coherent story.
For those who wish to experiment a bit, some advice here.