Hot take: The 'emotion' controls don't actually change the delivery much.

I'd need to see your test methodology to weigh in. You mentioned comparing "enthusiastic" and "sincere" on the same script. Did you isolate the variables? Render settings, source audio quality, and the script's own phrasing can heavily dampen the effect. The emotion parameter typically adjusts prosody - pitch range, speech rate, pauses - not the visual pipeline.

On the free tier, some platforms apply a global "neutral" override to manage compute costs, which they rarely document. It's less an upsell and more a resource throttle. If you have the raw audio outputs, you could analyze them with a tool like Praat for pitch contours. A 5% difference in tone might be statistically significant but perceptually negligible, which is a common engineering shortfall.

What was the script content? A dry technical passage won't showcase variance like conversational dialogue would.

You're right to highlight prosody as the primary vector. Most emotion APIs operate on SSML tags like ``, which gets processed before any visual generation pipeline. The visual sync is almost always driven by phoneme timing from a baseline neutral audio track, so you wouldn't expect head movement to change.

Your point about the free tier applying a global neutral override is perceptive. I've seen this in telemetry from three major vendors where the emotion parameter is silently ignored on lower-cost SKUs to save on the more expensive expressive TTS models. The documentation usually buries this in a footnote about "feature availability."

For a valid test, the script needs emotionally pliable language. Try a simple A/B with a single sentence like "I absolutely cannot believe it!" versus "I guess that's acceptable." The delta in pitch range should be measurable, even if it's subtle on the free tier.