Regularizing Learnable Feature Extraction for Automatic Speech Recognition