Table 2:

Depression detection performance for various models and AO-SOTA baselines based on F1-AVG, F1(ND), F1(D), and Speaker ID accuracy using the DAIC-WoZ dataset. SOTA baseline results are either reproduced values or reported from the corresponding study. The symbol ‘–’ indicates that those values were not reported in the corresponding study. The symbols ‘ ↑’ and ‘ ↓’ indicate a higher or lower value is better, respectively. Best results are highlighted in bold.

Model Architecture	Input Feature	Disentanglement Method	Model Parameters	F1-AVG ↑	F1(ND) ↑	F1(D) ↑	SID Accuracy ↓
DepAudioNet [34]	Mel-Spectrogram	None	280k	0.6081	0.6977	0.5185	-
FVTC-CNN [40]	Formants	None	-	0.6400	0.4600	0.8200	-
Speech SimCLR [41]	Mel-Spectrogram	None	-	0.6578	0.7556	0.5600	-
CPC [42]	Mel-Spectrogram	None	-	0.6762	0.7317	0.6207	-
CNN-LSTM [23]	Spk. Embd. + OpenSmile	None	-	0.6850	0.8600	0.5100	-
SpeechFormer [43]	Wav2Vec	None	33M	0.6940	-	-	-
Vowel-based [44]	Mel-Spectrogram	None	-	0.7000	0.8400	0.5600	-
DepAudioNet [37] (D1)	Raw-Audio	None	445k	0.6259	0.7755	0.4762	10.04%
DepAudioNet [30] (D2)	Raw-Audio	USD	459k	0.6830	0.7826	0.5833	8.91%
DepAudioNet (D3)	Raw-Audio	NUSD	459k	0.7086	0.8085	0.6087	8.05%
ECAPA-TDNN (E1)	Raw-Audio	None	595k	0.6329	0.7273	0.5385	42.33%
ECAPA-TDNN (E2)	Raw-Audio	USD	609k	0.7086	0.8085	0.6087	9.38%
ECAPA-TDNN (E3)	Raw-Audio	NUSD	609k	0.7349	0.8333	0.6364	4.68%
Δ (E3 vs E2) in %	-	-	-	3.70	2.80	4.55	−50.11