Table 4.

A comprehensive comparison of various deep learning-based architectures was conducted to assess their performance on the MAD dataset for audio classification.

architecture	parameters	pretrain	accuracy (%)
			pretrained weights	scratch
ResNet18	11.7M		88.04±0.59	86.83±0.38
ResNet34	21.8M	ImageNet	88.27±0.29	87.07±0.74
ResNet50	25.6M		88.24±0.41	86.51±0.38
ResNet101	44.7M		88.62±0.31	86.86±0.60
EfficientNet-B0	5.3M		88.92±0.17	87.32±1.00
EfficientNet-B1	7.8M	ImageNet	88.85±0.28	87.71±0.33
EfficientNet-B2	9.2M		89.11±0.64	88.02±0.27
CNN6	4.8M		89.53±0.26	87.57±0.81
CNN10	5.2M	AudioSet	90.76±0.47	${\underset{\_}{89.03}}_{\pm 0.37}$
CNN14	80.7M		${\underset{\_}{90.97}}_{\pm 0.34}$	89.47±0.22
AST	87.7M	ImageNet + AudioSet	90.03±0.23	81.17±0.77
AST-Patch-Mix		ImageNet + AudioSet	91.07±0.19	82.20±1.13

Employing pretrained model weights from the ImageNet⁴² and AudioSet²⁹ datasets is denoted as pretrained weights, while not using pretrained weights is denoted as scratch, respectively. Best and second best results.