For the training we split the set of recorded utterances into a training set and a
test set. These sets were formed separately for the two recognition tasks.
Yes/no-questions have a significant different prosody compared to normal
statements. Other types of questions do not necessarily use the same prosody. From
the 40 questions in our corpus only 20 are yes/no-questions. We split this set into 14
questions for training and six questions for testing. From the remaining utterances
we selected 16 sentences for the training. For testing we used the six yes/no
questions and eight sentences from the normal utterances. Using only the last word
of each utterance, this results in 30 words for training (14 questions/16 normal) and ...