features also improved performance signicantly. Linguistic
models were trained on IEMOCAP data, half of which was
selected from the spontaneous sessions. However, because
half of the utterances were from the scripted sessions, similar
lines of dialogue were likely present in both training and
test sets.