The quality layer for AI audio drama.
It listens back to its own render.
It reads your story like a listener, performs it like a full cast, and checks every line before the world hears it.
You know me, then?
directed surprise, heard neutral
Both paths heard it flat and agreed with each other, so the line does not ship. One path agreeing with the direction would have cleared it.
Every stage hands the next something it can check.
everything below ships
Story Analysis
hands over a cliffhanger band and the weakest beat
Cliffhanger band
A relative band with the rationale and the sources behind it, never a score.
One-click rewrite
The weakest beat comes back stronger, labelled demo or live.
Emotional arc
Named from six shapes, with the reference curve drawn over your own.
- Stability band
- Second-episode hook
- Foreshadowing
The Director
hands over one direction per line
A cast per character
With the reasoning for each voice, shown rather than hidden.
Emotion and delivery
Every line gets a directed emotion and a delivery note.
SFX spotted
Cues read out of the prose and matched to a bed and SFX bank.
- Music bed by mood
- Ducked under dialogue
- Voice lock
Audio QC
hands over publish, or hold
Two emotion paths
An acoustic read and an audio LLM, run independently on every line.
Held only when both disagree
One path agreeing with the direction is enough to clear the line.
The acoustic leg
A local SER model where the weights exist, a prosody proxy otherwise.
- The audio judge
- Every reading records its rung
- Transcript diff
- Pace
- Loudness
- Voice distinctiveness
- Voice consistency
- Pronunciation
- Regenerate one line
Ship it
hands the listener a link
The held take against the fix
Same line, same voice, side by side. Only the direction changed.
A share page
Publish to a link that plays anywhere. No account needed.
A generated cover
Typographic artwork built for the scene from its own mood.
- MP3 export
- A public feed
- Appended to the flywheel
above stage 3
The Producer can overrule the rule. It cannot hear the audio.
- Reads the measurements. Never the audio.
- Never replaces the rule's verdict. Both are returned.
- Its confidence is self-reported and uncalibrated.
- Unknown stakes block. It does not guess.
around the stages
The rest of the workspace.
A panel, not one verdict
Listener personas read the chapter in parallel and are allowed to disagree. Where they split is the finding: agreement tells a writer nothing they did not already know.
Meet the panelA report card, not a wall
Emotion, pronunciation, continuity, pacing and loudness each resolve to one row with its own findings, above the Producer's release decision.
Read the rows
measured, not claimed
The classifiers are scored, and reported as agreement.
Scored against a hand-labeled public-domain set: agreement with human labels, never a claim about absolute quality. The emotion labels are written affect rather than recorded speech, and the story signals stay relative bands with rationale.
| classifier | accuracy | cohen kappa | macro-F1 |
|---|---|---|---|
| Emotionn=30 | 0.867 | 0.843 | 0.882 |
| Story arcn=14 | 0.857 | 0.829 | 0.781 |