Towards Robust Waveform-Based Acoustic Models