Can you build an algorithm that understands simple speech commands?
My attempt at Kaggle's TensorFlow speech recognition challenge
- Convolution layers: 3x3 kernels with valid padding
- Max pool layers: 2x2 with stride 2
- First convolution layer: 32 filters
- Subsequent convolution layers double number of filters
- Dropout layer: trained at 0.4
- First fully connected layer: 128 units
- Second fully connected layer: 64 units
| layer | output shape | parameters |
|---|---|---|
| conv1 | (97, 159, 32) | 320 |
| pool1 | (97, 159, 32) | 0 |
| conv2 | (46, 77, 64) | 576 |
| pool2 | (23, 38, 64) | 0 |
| conv3 | (21, 36, 128) | 1,152 |
| pool3 | (10, 18, 128) | 0 |
| conv4 | (8, 16, 512) | 4,608 |
| pool4 | (4, 8, 512) | 0 |
| flat | (16384) | 0 |
| dense4 | (128) | 2,097,280 |
| dense5 | (64) | 8,256 |
| softmax | (10) | 650 |
| TOTAL | 2,112,842 |
