How to produce .npy features correctly in sound_classification_ml_production?

Question

I ran the github repo code sound_classification_ml_production and got good accuracy, but found that .npy features are pre produced by author. So I want produce my own .npy and a