Revisiting Interpolation Augmentation for Speech-to-Text Generation