Table 3.

Comparison of recognition rates among different methods for SAVEE dataset.

Method	Audio Representation	Accuracy
SVM [9]	feature vector †	77.4%
SVM [29]	feature vector †	48.81%
Random Forest [29]	feature vector †	56.07%
2D CNN [7]	Spectrograms	73.6%
VGG-16${}^{\left(1\right)}$	all spectrograms *	49.20%
VGG-16${}^{\left(2\right)}$	k selected spectrograms ⋄	45.11%
Proposed 3D CNN${}^{\left(1\right)}$	all spectrograms *	80.41%
Proposed 3D CNN${}^{\left(2\right)}$	k selected spectrograms ⋄	81.05%

^†: A feature vector of commonly known audio features like Table 1. ^*: All generated frames/spectrograms of one audio is used. ^⋄: Only k (9) frames/spectrograms of one audio is used.