Of course we need more test utterances to draw any reasonable conclusion, and probably longer utterances. Another way to measure the differences is to get the log-likelihood score of each utterance. Basically can give you an idea of which the model trained from these audio source fits better with test data. Normally we don't want any compression on our data, by heart, I would think the orignal wav will perform well. However, there is a possibility the compression helps to enhance the speech.