Feature-wise CCC rates for top-performing handcrafted Feature Representations (FR).

Code coming soon!

VAM database

Dimension	Type	FR ID	FR desription	CCC
Valence	Sysytem	mfcc_sma[1]_upleveltime75	segment duration where signal is above 0.75*range of 1st MFCC coefficient	0.199
Valence	System	mfcc_sma[3]_percentile1.0	first percentile of 3rd MFCC coefficient	0.171
Arousal	System	audspec_lengthL1norm_sma_quartile3	3rd quantile for magnitude of L1 norm of Auditory Spectrum	0.725
Arousal	System	audspec_lengthL1norm_sma_peakMeanAbs	arithmetic mean of peaks for L1 norm of Auditory Spectrum	0.716
Dominance	System	audspec_lengthL1norm_sma_percentile99.0	99th percentile for L1 norm of Auditory Spectrum	0.659
Dominance	System	audspec_lengthL1norm_sma_peakMeanAbs	arithmetic mean of peaks for L1 norm of Auditory Spectrum	0.669

IEMOCAP database

Dimension	Type	FR ID	FR desription	CCC
Valence	Sysytem	audspec_lengthL1norm_sma_de_flatness	contour flatness for delta values of magnitude of L1 norm of Auditory Spectrum	0.166
Valence	Sysytem	audSpec_Rfilt_sma_de[23]_flatness	ontour flatness for delta coefficient of 23rd coefficient of Relative Spectral Transform (RASTA)-style filtered applied to Auditory Spectrum	0.144
Arousal	Sysytem	mfcc_sma[2]_range	range of 2nd MFCC coefficient	0.469
Arousal	Sysytem	mfcc_sma[2]_pctlrange0-1	he range of the 1% and the 99% for 2nd MFCC coefficient	0.473
Dominance	Sysytem	mfcc_sma[2]_pctlrange0-1	he range of the 1% and the 99% for 2nd MFCC coefficient	0.375
Dominance	Sysytem	mfcc_sma_de[4]_lpc1	1st linear prediction filter coefficient for the delta coefficient of 4th MFCC coefficient	0.355

For both datasets, the top feature-wise CCC rates. FR represents vocal tract systems. Descriptions of the top-performing FR can be found in the four columns of tables.

The Python code for extracting WavLM FR used in our study can be found in the following repository: https://github.com/idiap/ExVo-2022