PhD A thought the results showed that delta features were important and that the TIMIT-trained network performed best overall, though it was unclear why. They found that neural networks helped when trained on the target task or when feature normalization was poor, but could hurt performance on a different task; MSG features also did not improve the Italian results.