Robustness spectral tilt

Overall scores

	w2v2-L	hubert-L	wavlm	data2vec
Overall Score	62.5% passed tests (5 passed / 3 failed).	87.5% passed tests (7 passed / 1 failed).	75.0% passed tests (6 passed / 2 failed).	62.5% passed tests (5 passed / 3 failed).

Change Ccc Downward Tilt

Threshold: -0.05

Data	Change CCC Downward Tilt
	w2v2-L	hubert-L	wavlm	data2vec
iemocap-2.3.0-emotion.dimensions.test.gold_standard	0.00	-0.01	-0.03	0.00
msppodcast-2.6.1-emotion.dimensions.test-1.gold_standard	0.00	-0.01	-0.03	-0.01
mean	0.00	-0.01	-0.03	-0.01

Change Ccc Upward Tilt

Threshold: -0.05

Data	Change CCC Upward Tilt
	w2v2-L	hubert-L	wavlm	data2vec
iemocap-2.3.0-emotion.dimensions.test.gold_standard	-0.06	-0.03	0.04	-0.07
msppodcast-2.6.1-emotion.dimensions.test-1.gold_standard	-0.03	-0.01	-0.01	-0.04
mean	-0.04	-0.02	0.01	-0.06

Percentage Unchanged Predictions Downward Tilt

Threshold: 0.8

Data	Percentage Unchanged Predictions Downward Tilt
	w2v2-L	hubert-L	wavlm	data2vec
iemocap-2.3.0-emotion.dimensions.test.gold_standard	0.98	1.00	0.99	0.94
msppodcast-2.6.1-emotion.dimensions.test-1.gold_standard	0.97	0.99	0.89	0.85
mean	0.97	0.99	0.94	0.90

Percentage Unchanged Predictions Upward Tilt

Threshold: 0.8

Data	Percentage Unchanged Predictions Upward Tilt
	w2v2-L	hubert-L	wavlm	data2vec
iemocap-2.3.0-emotion.dimensions.test.gold_standard	0.42	0.77	0.59	0.38
msppodcast-2.6.1-emotion.dimensions.test-1.gold_standard	0.75	0.85	0.72	0.54
mean	0.58	0.81	0.66	0.46

Visualization Downward Tilt

Difference of predictions for original audio and audio with a downward spectral tilt. The allowed prediction difference \(\delta < 0.05\) is highlighted in green in the upper plot. The lower plot shows the distributions of the two predictions.

w2v2-L	hubert-L	wavlm	data2vec