Table 1.

CNN architecture used for the sound event and speech command datasets.

	Sound Event	Speech Command
Image input layer	32 × 15	64 × 64
Middle layers	Conv. 1: 16@3 × 3, Stride 1 × 1, Pad 1 × 1 Batch Normalization, ReLU Max Pool: 2 × 2, Stride 1 × 1, Pad 1 × 1 Conv. 2: 16@3 × 3, Stride 1 × 1, Pad 1 × 1 Batch Normalization, ReLU Max Pool: 2 × 2, Stride 1 × 1, Pad 1 × 1	Conv. 1: 48@3 × 3, Stride 1 × 1, Pad ‘same’ Batch Normalization, ReLU Max Pool: 3 × 3, Stride 2 × 2, Pad ‘same’ Conv. 2: 96@3 × 3, Stride 1 × 1, Pad ‘same’ Batch Normalization, ReLU Max Pool: 3 × 3, Stride 2 × 2, Pad ‘same’ Conv. 3: 192@3 × 3, Stride 1 × 1, Pad ‘same’ Batch Normalization, ReLU Max Pool: 3 × 3, Stride 2 × 2, Pad ‘same’ Conv. 4: 192@3 × 3, Stride 1 × 1, Pad ‘same’ Batch Normalization, ReLU Conv. 5: 192@3 × 3, Stride 1 × 1, Pad ‘same’ Batch Normalization, ReLU Max Pool: 3 × 3, Stride 2 × 2, Pad ‘same’ Dropout: 0.2
Final layers	Fully connected layer: 50 Softmax layer Classification layer	Fully connected layer: 36 Softmax layer Classification layer