Why we need evals for audio generative AI models?

What criteria should we use to determine the best model when we have text-to-speech models such as ElevenLabs, Bark, etc?

How do we scale this up when these audio models have their "stable diffusion moment" (thanks simonw for the phrase).