Robustifying automatic speech recognition by extracting slowly varying features