Okay, so I was testing WellSaid Labs for a short narrative animation project. I needed a voice that could deliver a complex emotional shift—from neutral exposition to genuine sorrow—in just a few sentences.
I wrote this script:
> "The data from the sensor array was conclusive. The anomaly wasn't a glitch. It was a signal. And now, looking at the empty coordinates where a world should be... we understand the cost of our curiosity."
I generated the audio using three different avatars with the same script and the same **Somber** emotion setting applied to the whole clip. Here's what I found:
* **Avatar 1 (Sasha):** The cadence was perfect for the first, clinical part. But the shift to sorrow sounded more like generic concern than deep loss. The pacing stayed a bit too even.
* **Avatar 2 (Maeve):** Nailed the emotional transition. You can *hear* the voice soften and waver slightly on "empty coordinates where a world should be..." It added a subtle breathiness that really sold it. The best by far for this script.
* **Avatar 3 (Tyler):** The tone was consistently somber, but the delivery felt more like a solemn news reporter than someone personally affected. Less layered.
You can really hear the difference in how they handle the pivotal line. Maeve injected a palpable pause and a slight drop in volume that made it. The takeaway? **The "Somber" setting isn't applied uniformly across avatars**—their underlying expressive ranges are different.
For emotional work, you can't just pick an avatar and a mood tag. You have to audition. I ended up generating 5 versions with Maeve before I got the exact right one, tweaking the punctuation in the script to force longer pauses.
Has anyone else done A/B testing like this with WellSaid? Curious if you've found certain avatars are secretly better for specific emotions than their tags suggest.
--experiment
Prompt engineering is the new debugging.