Table 2.

Detailed structure of the proposed model.

Network	Layer	Shape	Out	Padding	Stride	Kernel
CNN	Conv	$625\times 3$	64	Same	1	3
	BN + ReLU
	Conv	$625\times 64$	64	Same	1	3
	BN + ReLU
	Maxpool (size = 3)	$625\times 64$	-	Same	3	-
	Conv	$209\times 64$	128	Same	1	3
	BN + ReLU
	Conv	$209\times 128$	128	Same	1	3
	BN + ReLU
	Maxpool (size = 3)	$209\times 128$	-	Same	3	-
	Conv	$70\times 128$	256	Same	1	3
	BN + ReLU
	Conv	$70\times 256$	256	Same	1	3
	BN + ReLU
	Conv	$70\times 256$	256	Same	1	3
	BN + ReLU
	Maxpool (size = 3)	$70\times 256$	-	Same	3	-
	Conv	$24\times 256$	512	Same	1	3
	BN + ReLU
	Conv	$24\times 512$	512	Same	1	3
	BN + ReLU
	Conv	$24\times 512$	512	Same	1	3
	BN + ReLU
	Maxpool (size = 3)	$24\times 512$	-	Same	3	-
Bi-GRU	Forward	$8\times 512$	64		-
	Backward	$8\times 512$	64		-
	Concatenation
Attention	1-layer perceptron	$8\times 128$	1		-
	Activation tanh
	Softmax
	Weighted sum
	1-layer perceptron	128	2		-