Eleven v4: direct a performance and compare it fairly

Eleven v4 accepts natural-language performance directions and inline audio tags. That makes a useful test possible: keep the sentence and voice fixed, change only the direction, and listen for the intended difference.

I would start with one sentence that can carry two meanings: “That was a brilliant idea.” One version should sound sincerely reassuring; the other should sound dry and sarcastic. Neither performance is automatically better. The question is whether the model follows your intended delivery while keeping the words understandable.