Table 2:

F1-scores for DepAudioNet using the CONVERGE dataset for three features, with and without adversarial speaker disentanglement. ‘+Adv’ denotes adversarial training.

Input Feature	+Adv	F1-Avg	F1-ND	F1-D
Mel-spec	No	0.879	0.890	0.868
Mel-spec	Yes (λ = 5e-5)	0.890	0.903	0.877
Raw-audio	No	0.829	0.832	0.826
Raw-audio	Yes (λ = 2e-4)	0.857	0.870	0.844
XLSR-Mandarin	No	0.912	0.921	0.903
XLSR-Mandarin	Yes (λ = 2e-4)	0.915	0.925	0.906