isn't that supposed to be the magic of convnets though? they _figure out_ the right format. instead of doing feature engineering, like stfts and mel warping,

isn't that supposed to be the magic of convnets though? they _figure out_ the right format. instead of doing feature engineering, like stfts and mel warping, you do stuff like build convolution layers into an ann and let it sort it out?