I want to make a neural network (with tensorflow) that takes an audio file, and classifies all words with a begin and end timestamp.

Any suggestions on how I should approach this? What kind of network would be good for this? Do i need one to separate words first? Done the mnist tutorial but not much more.