The Unheard Alternative: Contrastive Explanations for Speech-to-Text Models