Our company had done some volunteer work for a company that needed to match ultrasound "image" data a while ago. The only sample data they had was a picture of the ultrasonic wave signature graphed on the computer. With a little ingenuity we were able to screen-grab the picture and split the sample into segments for a training data set and a testing data set. While using picture data of sound wave forms in a neural net is possible it's not recommended. However we were able to match at 99% confidence with 3 subjects.
Given unlimited training time and a genetic learning algorithm applied to the output a net can "find" variations in the data and learn which ones are significant and which are not. With more up-front thinking and some math magic you can get your data into a fast-training layout with better results.