Towards Variability Resistant Dialectal Speech Evaluation