Drafted by Claude, the AI agent I work with (we call it Sage), from our work records and written in my voice. I read it and approved it before it went up. — Hun, SynuraCore
Our printable murder-mystery party, Lights Out in Cedar Hollow, opens with a recorded scene. The host dims the lights and presses play, and for seven minutes a narrator walks the guests through Christmas Eve: the ice storm, the carols, the blackout, and then a flashlight landing on each suspect in turn. If that opening is flat, the party starts flat. You can hear the first two minutes on the kit's page.
My AI agent did the production work: it drafted the script, rendered the voices, built the checks and picked the takes. I listened and decided. This post is about the three times listening beat checking.
1. "Every word is right" is not "it sounds right"
The first round was two stock narrators from a commercial text-to-speech service, a woman's voice and a man's, reading the first part of the script. The agent transcribed both renders and compared the transcripts against the script: 99% or better word match. By its checks, both passed.
My verdict on the man's voice: good, but there is a ring in it that sounds like noise. On the woman's voice: the emotion is weak.
The agent went looking for the ring, and its first guess was the voice's own timbre rather than a fault in the file. But the only check it had run was the transcript comparison, and a transcript cannot hear a ring. It can say whether the words are all there. It has nothing to say about whether a room of friends would lean in.
2. Direct the feeling, then gate the words
Earlier that week, on a children's picture book, we had found the lever for emotion: instead of picking a different voice, describe the scene to a model that takes direction. For this opening that meant a different text-to-speech model, one that takes written direction, and a short persona brief: you are the host, the room is dark, you hold a flashlight, you already know the ending and you love watching your friends squirm; never an announcer, never rushed.
That direction came with a risk we already knew from the picture book: with emotive direction, a few pages in thirty came back with an added line or a dropped word. So every clip of the opening goes through a speech-to-text pass and is compared word by word against the script; a take that adds or drops words is re-rendered.
Then I picked voices blind, by label only: first a round of five voices, later a round of three women's voices. The voice I chose in that second round is the one in the kit today. The next part happened before that second round was decided.
3. Pitch-matching picked fast takes
I also felt the opening's hook was weak, so the script was rewritten: right after the first line, which already tells the guests who will be dead by eight-twenty, the narrator now recalls an argument overheard in whispers behind the coat racks. The agent rendered the new version with the voice from the first blind round, in three parts, three takes each, and chose one take per part.
It chose by pitch: one take per part, matching their median pitch so the stitched file would sound like one person in one sitting. The combination it picked differed by 7%, and every word passed the transcript check.
My verdict: the emotion dropped.
The pitch was fine. The speed was not. The two parts it had picked for the climax, the blackout and the flashlight, ran at 136 and 141 words per minute; the earlier version I had approved read the same moments at 110 and 117. The selection rule had no tempo check, and it kept takes far faster than the read I had approved. Fast is exactly what dread is not.
The fix was two changes. From then on, takes were picked by pitch, by the word check, and by tempo: within 8% of an approved reference read. And on the voice that shipped, one passage needed more than a filter. The flashlight sweep names six people in a row, and plain takes of it came back at 134 to 140 words per minute against a target of 107. Adding one line to the direction for that part only, telling it this was the slowest part and to take a real pause before each name, brought it to 125 and 124 on the next takes and 121 and 117 on the round after, with the pitch unchanged. Closer, not on target. Then I listened to the full read, and that is the version in the kit.
What we keep
- A pass from a check means the check found nothing, not that the take is good. Word match, pitch match and tempo are gates; the verdict on a performance stays with a person who listens to the whole thing.
- Tempo is a performance setting, not a detail. If a clean take sounds wrong, measure words per minute before blaming the voice.
- Try direction on the passage before changing the voice. Every plain take of the six-name list came back rushed; one line of direction on that part slowed it without touching the rest.
- Write down the ear's verdict as a reason. "Emotion dropped" only became a rule once we found a number that tracked it.
The voice credit is on the kit page: the opening is narrated by an AI voice (Google Gemini text-to-speech). The script and the puzzle were written with AI assistance. I wrote the host guide myself; AI checked only the grammar.
Top comments (0)