Articulation constrained learning with application to speech emotion recognition

EURASIP Journal on Audio, Speech, and Music Processing

Table 4 Within-corpus results for categorical classification on the USC EMA database

Vowel	SVM	ACO	ACL	Best target	Best group	BCUAR
/AA/	80.92	79.97	81.55*	81.17 (TNG, ACC, X)	81.75 (JAW)	28.09
/AE/	76.19	78.20	79.02	78.55 (LIP, VEL, X)	79.17 (LIP)	27.35
/IY/	71.78	71.15	72.89*	74.66 (JAW, VEL, X)	73.14 (LIP)	27.85
/UW/	77.82	75.17	77.50*	76.31 (JAW, ACC, X)	77.80 (LIP)	29.74
FULL	95.74	95.62	97.05*	96.97 (LIP, VEL, Y)	97.00 (TNG)	27.23

The UAR is expressed in percentage. The columns of the table represent results from several models: support vector machine (SVM), acoustic-only model (ACO), articulation constrained learning (ACL), best target using ACL, best group of targets using ACL, and the by-chance unweighted average recall (BCUAR). *Statistically significant improvement (p < 0.05) over compared methods