. Author manuscript; available in PMC: 2016 Oct 1.

Published in final edited form as: Comput Speech Lang. 2015 Jan;29(1):132–144. doi: 10.1016/j.csl.2014.02.001

Table 5.

Results show the unweighted (weighted) average recall of I and NI labels on test set of sub-challenge dataset before and after posterior smoothing. The hyperparameter σ was learned on the development set. ‘All’ is feature-level fusion with all subsystem’ features. All accuracies in %.

	KNN		SVM		LDA

Posterior smoothing	Before	After	Before	After	Before	After

Prosody	66.3(64.8)	69.5(73.2)	64.9(65.2)	71.3(75.5)	66.0(69.1)	65.3(69.1)
Pronunciation	66.7(65.6)	71.0(71.7)	67.5(65.9)	71.7(71.3)	52.9(50.0)	56.3(53.0)
Voice quality	59.2(60.7)	65.9(68.0)	64.7(64.8)	66.3(66.0)	62.0(62.2)	68.9(71.7)
All	66.1(65.0)	72.0(73.9)	69.6(71.1)	69.3(72.3)	67.9(70.7)	67.2(70.7)