Table 3.

Classification results of feature selection and parameter tuning on the pathology challenge data. ‘Dim.’ is the feature dimension after forward feature selection. ‘Acc.’ is unweighted (weighted) average recall in % on test set using I-best features which are chosen based on their classification accuracy on the development set. By-chance is 50% for unweighted average recall and 64.4% for weighted average recall. Note that we evaluated performance based on unweighted average recall to be consistent with the criterion of the IS2012 pathology sub-challenge. ‘All’ is feature-level fusion with all subsystems’ features.

	LDA		KNN			SVM
Feature set	Dim.	Acc.	k	Dim.	Acc.	Kernel	Dim.	Acc.
Prosody	6	66.0(69.1)	15	6	66.3(64.8)	2nd-order poly.	7	64.9(65.2)
Pronunciation	2	52.9(50.0)	19	5	66.7(65.6)	3rd-order poly.	6	67.5(65.9)
Voice quality	5	62.0(62.2)	14	6	59.2(60.7)	2nd-order poly.	5	64.7(64.8)
All	11	67.9(70.7)	15	10	66.1(65.0)	2nd-order poly.	12	69.6(71.1)