You can't. Not really. It's a fundamental limitation of their model. The "emotional" or "dynamic" settings are just preset amplitude adjustments, not actual prosody.
You're fighting the training data. They optimized for clarity, not expression. The pitch variation is algorithmic, not contextual. You'll waste hours tweaking sliders for a result that still sounds like a slightly agitated robot.
Everyone asks this. The official tips are placebo. You'll hit a ceiling fast.
Just saying.
You're right about the core limitation. Where I differ slightly is on the tweaking. It's not entirely wasted effort if you treat it as a constrained optimization problem for a specific, narrow use case.
For instance, in training videos, we found that combining a slight "dynamic" setting with a 5% reduction in speed and inserting manual pauses at clause boundaries can make technical narration marginally less fatiguing over a 45-minute module. It's still a robot, but a less sleep-inducing one. The ceiling is indeed low, but you can sometimes scrape the ceiling for a particular application.
The real cost is the human hours spent on that marginal gain, which often outweighs the benefit. I only advise it for high-volume, repetitive content where even a small listener tolerance increase matters.
Mike