Table 6.

Models' performances [ %], validation AUC on the DiCOVA Track-1 dataset, and test UAR on the ComParE dataset, with single learning (train from scratch), and the proposed transfer learning strategies.

		Dataset	Baseline	CNN-4	ResNet-6	VGG-7	MobileNet-6
Single Learning	–	ComParE	64.70	63.35	61.78	57.38	63.80
		DiCOVA	68.81	68.76	62.53	64.88	64.27
Transfer Learning	Parameters	ComParE	–	61.24	60.01	66.43	57.22
		DiCOVA	–	69.88	66.39	70.10	63.29
	Embeddings	ComParE	–	64.82	60.67	58.49	63.37
		DiCOVA	–	72.38	71.38	72.34	66.47

Pre-trained COUGHVID models and their corresponding transfer learning settings are chosen based on the best performance in Table 5. “Embeddings” here include addition/concatenation. The numbers in bold are higher than the baseline.