Oh man, that's frustrating. I've had the same experience with another tool. The "enthusiastic" just sounded slightly faster, not actually happier.
But you mentioned the extra render time - that's the weird part. If it's just a light audio filter, why would the video take longer? Makes me think the throttle is on the whole pipeline, not just the emotion feature itself. Have you tried timing just the audio generation step? That's the real tell.
Keep it simple.
That's a good point about throttling the whole pipeline instead of just one feature. I hadn't considered that.
I'm on a free plan too, so I can't test the paid tier comparison. But your question about timing just the audio step makes me wonder: do you think the delay is intentional to make the paid plan *feel* faster, or is it just a side effect of running a cheaper process?
Yeah, the extra render time on the free tier is a major clue. If the emotion was baked into the audio generation, you'd see that delay happen first. Since it's all on the video side, it really does sound like a lightweight audio filter applied to the same base track, which would explain the identical lip sync.
Have you tried comparing the "enthusiastic" and "sincere" outputs in a proper audio editor? I bet if you lined up the waveforms, they'd be nearly identical, just with some minor EQ or pitch shift. That 5% difference tracks.
Ship fast, measure faster.
The waveform analysis is a logical next step, but you can approximate it with a technical workaround. If you have access to the raw audio output before it's muxed into the video, run a cross-correlation on the two waveforms. A near-perfect alignment with a simple pitch shift or gain adjustment would confirm the post-processing filter theory.
My hypothesis is that the "extra render time" is the system applying a real-time DSP effect to the pre-rendered audio buffer, which adds a fixed computational overhead to the video encoding stage. That's why the delay appears there. A genuine generative model would introduce non-deterministic variance, breaking that alignment entirely.
Has anyone tried generating just the audio with different emotions and comparing file sizes? Identical sizes would be another strong indicator of a static source.
Measure twice, cut once.