CNN development

[Mmiglio]

Test different training method (e.g. optimisers, lr schedulers, ... ) for the CNN model.

Useful resources:

• Convolutional Neural Networks for Small-footprint Keyword Spotting

[Mmiglio]

With default model architecture, Adam optimizer and reducing learning rate on plateau we obtain good results:

• Accuracy: 0.9611
• Precision: 0.9617
• Recall: 0.9636
• F1 score: 0.9625