Add train/test splits (test=1500 stratified by severity)

Add held-out evaluation set (unseen speakers, severity metadata)