[Feature request] Add evaluator for speech2text models

Now that Whisper is integrated into `transformers` and the large number of speech2text models on the Hub, it might be interesting having a speech2text evaluator that computes common metrics like word error rate etc.

This would be especially useful for `optimum` users, who want to know if their exported models suffer any degradation from quantization etc.

I don't have bandwidth to tackle this right now, but maybe someone in the community is excited to?

cc @mht-sharma