What this shows (for dummies). For 12 diverse prompts (all five DramaBox subsets, two-scene and single-emotion), we grew the take pool from 16 to 128 (with_dialogue instruction, no reference audio). For each of the three rewards you see the best take the original group of 16 could offer next to the top-3 of all 128 — listen to what 8× more sampling actually buys.
| reward | best-of-16 | best-of-128 | Δ | gain % | groups improved |
|---|---|---|---|---|---|
| all: invWER × (norm prompt-sim + blend + genu) | 1.473 | 1.672 | 0.199 | 13.5% | 83% |
| original: invWER × (norm blend + genu) | 0.950 | 1.105 | 0.154 | 16.3% | 92% |
| adherence: prompt-sim × invWER | 0.075 | 0.121 | 0.046 | — | 75% |
Full DramaBox prompt (how to say — with the "quoted" dialogue):
A 38-year-old woman with a warm but weary timbre, delivering this high-quality studio voice recording with no background noise.
The same voice is struggling to hold onto a practiced sadness, but the familiar silence is making her unexpectedly restless.
(letting out a slow breath) "I thought I'd feel... more. More tears, more ache." (pausing, listening to the creak) "Just dust motes in the sun."
CUT TO:
The same voice now lets out a surprised, private laugh as the emptiness finally feels like permission.
(shaking her head, almost amused) "God, the quiet isn't heavy. It's just... air." (hearing her own footsteps) "I can walk loud now. No one to hear."
The performance across both moments should feel like a single person slowly untangling grief from relief, each breath a small discovery.
Direction only (dialogue removed — this is what prompt-sim compares against):
A 38-year-old woman with a warm but weary timbre, delivering this high-quality studio voice recording with no background noise.
The same voice is struggling to hold onto a practiced sadness, but the familiar silence is making her unexpectedly restless.
(letting out a slow breath) (pausing, listening to the creak)
CUT TO:
The same voice now lets out a surprised, private laugh as the emptiness finally feels like permission.
(shaking her head, almost amused) (hearing her own footsteps)
The performance across both moments should feel like a single person slowly untangling grief from relief, each breath a small discovery.
What to say (goes in MOSS text field): I thought I'd feel... more. More tears, more ache. Just dust motes in the sun. God, the quiet isn't heavy. It's just... air. I can walk loud now. No one to hear.
Full DramaBox prompt (how to say — with the "quoted" dialogue):
A 34-year-old woman with a voice that is clear and measured, carrying a faint tremor of old hurt, delivering this high-quality studio voice recording with no background noise.
The same voice is tightly controlled, as if she has rehearsed these words for years, but the stillness is brittle.
(quiet, deliberate) "I was seven. I sat by the window every day after school." (a slow, measured exhale) "I thought you'd just been delayed."
CUT TO:
The same voice now cracks open into raw, childlike devastation, the adult composure gone.
(voice breaking, barely a whisper) "If I forgive you… does that mean what you did was okay?" (a long, shuddering pause) "Was I that difficult to love?"
The performance across both moments should feel like a dam breaking—from a rigid facade of composure to the naked, unfiltered ache of a child who never stopped waiting.
Direction only (dialogue removed — this is what prompt-sim compares against):
A 34-year-old woman with a voice that is clear and measured, carrying a faint tremor of old hurt, delivering this high-quality studio voice recording with no background noise.
The same voice is tightly controlled, as if she has rehearsed these words for years, but the stillness is brittle.
(quiet, deliberate) (a slow, measured exhale)
CUT TO:
The same voice now cracks open into raw, childlike devastation, the adult composure gone.
(voice breaking, barely a whisper) (a long, shuddering pause)
The performance across both moments should feel like a dam breaking—from a rigid facade of composure to the naked, unfiltered ache of a child who never stopped waiting.
What to say (goes in MOSS text field): I was seven. I sat by the window every day after school. I thought you'd just been delayed. If I forgive you… does that mean what you did was okay? Was I that difficult to love?
Full DramaBox prompt (how to say — with the "quoted" dialogue):
A 34-year-old woman with a tight, slightly clipped voice, delivering this high-quality studio voice recording with no background noise.
The same voice is bristling with controlled impatience, barely tolerating a delay in her morning rounds.
(tapping a pen against a clipboard) "This bloodwork should've been flagged two hours ago." (glancing at a chart) "Who signed off on this? Come on, people."
CUT TO:
The same voice now dissolves into raw grief, caught off guard by a patient's sudden death on the table.
(a long pause, then a whisper) "We did everything... everything right." (voice cracks) "Why does that never feel like enough?"
The performance across both moments should feel like a mask of professionalism cracking open into unguarded sorrow.
Direction only (dialogue removed — this is what prompt-sim compares against):
A 34-year-old woman with a tight, slightly clipped voice, delivering this high-quality studio voice recording with no background noise.
The same voice is bristling with controlled impatience, barely tolerating a delay in her morning rounds.
(tapping a pen against a clipboard) (glancing at a chart)
CUT TO:
The same voice now dissolves into raw grief, caught off guard by a patient's sudden death on the table.
(a long pause, then a whisper) (voice cracks)
The performance across both moments should feel like a mask of professionalism cracking open into unguarded sorrow.
What to say (goes in MOSS text field): This bloodwork should've been flagged two hours ago. Who signed off on this? Come on, people. We did everything... everything right. Why does that never feel like enough?
Full DramaBox prompt (how to say — with the "quoted" dialogue):
A 27-year-old female with a hushed, slightly breathy timbre, delivering this high-quality studio voice recording with no background noise.
The same voice is frozen in confusion and embarrassment, speaking with agonizing restraint as she tries to justify a mistake.
(voice barely above a whisper) "I just... I thought the chart was for room three." (swallows hard) "But the numbers didn't line up." (glances at the floor) "I'm sorry, I should have double-checked." (trails off) "I don't know what I was thinking."
CUT TO:
The same voice now sharp with indignation and sudden clarity, speaking with cold, restrained fury.
(stares directly ahead) "No, that's not acceptable. I need the real numbers now." (firmly) "You will explain this to me right now." (flat but intense) "I've had enough of the confusion."
The performance across both moments should feel like the same person transforming from paralyzed self-doubt to assertive clarity.
Direction only (dialogue removed — this is what prompt-sim compares against):
A 27-year-old female with a hushed, slightly breathy timbre, delivering this high-quality studio voice recording with no background noise.
The same voice is frozen in confusion and embarrassment, speaking with agonizing restraint as she tries to justify a mistake.
(voice barely above a whisper) (swallows hard) (glances at the floor) (trails off)
CUT TO:
The same voice now sharp with indignation and sudden clarity, speaking with cold, restrained fury.
(stares directly ahead) (firmly) (flat but intense)
The performance across both moments should feel like the same person transforming from paralyzed self-doubt to assertive clarity.
What to say (goes in MOSS text field): I just... I thought the chart was for room three. But the numbers didn't line up. I'm sorry, I should have double-checked. I don't know what I was thinking. No, that's not acceptable. I need the real numbers now. You will explain this to me right now. I've had enough of the confusion.
Full DramaBox prompt (how to say — with the "quoted" dialogue):
A late-middle-aged female with a grainy, chest-rich alto, delivering this high-quality studio voice recording with no background noise.
The same voice is simmering with sour jealousy, each word laced with tight-lipped envy.
(Bitterly.) "You know, she got the promotion just by being there. I've been here twice as long, but who counts that anymore." (Voice tightens.) "It's fine.
Direction only (dialogue removed — this is what prompt-sim compares against):
A late-middle-aged female with a grainy, chest-rich alto, delivering this high-quality studio voice recording with no background noise.
The same voice is simmering with sour jealousy, each word laced with tight-lipped envy.
(Bitterly.) (Voice tightens.) "It's fine.
What to say (goes in MOSS text field): You know, she got the promotion just by being there. I've been here twice as long, but who counts that anymore.
Full DramaBox prompt (how to say — with the "quoted" dialogue):
A very old man with a brisk, forward-pushing tempo, a low, resonant but tremulous and breathy voice, delivering this high-quality studio voice recording with no background noise.
The same voice is impatient and irritable but threaded with hope and affection, as if waiting for something anticipated.
"I swear, if that package doesn't show up today, I'll—but no, it's coming. They promised. And my granddaughter painted me a card, you know? She's seven. She uses too much glitter." (pause) "I'll keep it forever."
CUT TO:
The same voice now hollowed by grief, the hope drained away, speaking with a fragile, stunned quiet.
"It's just... the card. That's all I have now. The glitter's still on my fingers. She was supposed to be here. She was supposed to—" (weakly) "—be here."
The performance across both moments should feel like a single natural breath, from bustling expectancy to devastation, with the same old-man tremor and storybook cadence carrying through both.
Direction only (dialogue removed — this is what prompt-sim compares against):
A very old man with a brisk, forward-pushing tempo, a low, resonant but tremulous and breathy voice, delivering this high-quality studio voice recording with no background noise.
The same voice is impatient and irritable but threaded with hope and affection, as if waiting for something anticipated.
(pause)
CUT TO:
The same voice now hollowed by grief, the hope drained away, speaking with a fragile, stunned quiet.
(weakly)
The performance across both moments should feel like a single natural breath, from bustling expectancy to devastation, with the same old-man tremor and storybook cadence carrying through both.
What to say (goes in MOSS text field): I swear, if that package doesn't show up today, I'll—but no, it's coming. They promised. And my granddaughter painted me a card, you know? She's seven. She uses too much glitter. I'll keep it forever. It's just... the card. That's all I have now. The glitter's still on my fingers. She was supposed to be here. She was supposed to— —be here.
Full DramaBox prompt (how to say — with the "quoted" dialogue):
A mature androgynous adult with a deeply throaty, cavernous resonance and a deliberately labored, sedate tempo, delivering this high-quality studio voice recording with no background noise.
The same voice is intensely affectionate yet laced with contempt, a loving but scornful tone.
(Drawling, half-smile in voice.) "You know, I really do care about you. It's just that you make it so hard sometimes, the way you never listen, always so sure of yourself."
CUT TO:
The same voice now hollow with sudden
Direction only (dialogue removed — this is what prompt-sim compares against):
A mature androgynous adult with a deeply throaty, cavernous resonance and a deliberately labored, sedate tempo, delivering this high-quality studio voice recording with no background noise.
The same voice is intensely affectionate yet laced with contempt, a loving but scornful tone.
(Drawling, half-smile in voice.)
CUT TO:
The same voice now hollow with sudden
What to say (goes in MOSS text field): You know, I really do care about you. It's just that you make it so hard sometimes, the way you never listen, always so sure of yourself.
Full DramaBox prompt (how to say — with the "quoted" dialogue):
A young male with a deep, cartoonishly authoritative voice, tinny and bright yet booming, delivering this high-quality studio voice recording with no background noise.
The same voice is overwhelmed with intense gratitude, but the words sound annoyed.
(Sighs heavily) "You really didn't have to go and do all this." (Shakes head) "It's just
Direction only (dialogue removed — this is what prompt-sim compares against):
A young male with a deep, cartoonishly authoritative voice, tinny and bright yet booming, delivering this high-quality studio voice recording with no background noise.
The same voice is overwhelmed with intense gratitude, but the words sound annoyed.
(Sighs heavily) (Shakes head) "It's just
What to say (goes in MOSS text field): You really didn't have to go and do all this.
Full DramaBox prompt (how to say — with the "quoted" dialogue):
A 30-year-old woman with a warm, slightly husky alto, delivering this high-quality studio voice recording with no background noise.
The same voice is lying on cold concrete, eyes closed, feeling a single bare bulb's warmth on her face, savoring stillness before the door opens.
(soft whisper) "...warm light on my eyelids... I can almost feel the sun..." (barely audible, reverent) "this is my skin, this is my breath... I am still here..."
CUT TO:
The same voice now feels a metallic clamp on each ankle, then electricity seizing her body—the hum becomes a roar of fire, shattering her voice.
(jaw locked, teeth chattering) "No... NONONO—" (high-pitched, breathless shriek) "—it's in my BONES..." (last word dissolves into a sob that becomes a raw, rhythmic grunt)
The performance across both moments should feel like a woman being torn from a sacred, silent prayer into a brutal, all-consuming current.
Direction only (dialogue removed — this is what prompt-sim compares against):
A 30-year-old woman with a warm, slightly husky alto, delivering this high-quality studio voice recording with no background noise.
The same voice is lying on cold concrete, eyes closed, feeling a single bare bulb's warmth on her face, savoring stillness before the door opens.
(soft whisper) (barely audible, reverent)
CUT TO:
The same voice now feels a metallic clamp on each ankle, then electricity seizing her body—the hum becomes a roar of fire, shattering her voice.
(jaw locked, teeth chattering) (high-pitched, breathless shriek) (last word dissolves into a sob that becomes a raw, rhythmic grunt)
The performance across both moments should feel like a woman being torn from a sacred, silent prayer into a brutal, all-consuming current.
What to say (goes in MOSS text field): ...warm light on my eyelids... I can almost feel the sun... this is my skin, this is my breath... I am still here... No... NONONO— —it's in my BONES...
Full DramaBox prompt (how to say — with the "quoted" dialogue):
A 30-year-old woman with a husky, sand-warmed contralto, delivering this high-quality studio voice recording with no background noise.
The same voice is soft with melancholy, her palm pressed to cold rippled glass as she hums a slow, forgotten lullaby.
(whispering near the glass) "This lock's been dead for years... the hinges are the only things that remember how to sing." (letting the thought dissolve into a breathy sigh)
CUT TO:
The same voice now yanked into her throat by a sudden face on the glass, her breath stolen, then ripped out as a raw, high-velocity scream that splinters into a trembling, bitter laugh.
(stumbling backward, voice cracking) "Oh, you think that's funny, do you? Just... just a little janitor?" (trailing into a jagged sob that she swallows hard)
The performance across both moments should feel like a solitary woman's intimate stillness shattered into a feral, wounded animal response.
Direction only (dialogue removed — this is what prompt-sim compares against):
A 30-year-old woman with a husky, sand-warmed contralto, delivering this high-quality studio voice recording with no background noise.
The same voice is soft with melancholy, her palm pressed to cold rippled glass as she hums a slow, forgotten lullaby.
(whispering near the glass) (letting the thought dissolve into a breathy sigh)
CUT TO:
The same voice now yanked into her throat by a sudden face on the glass, her breath stolen, then ripped out as a raw, high-velocity scream that splinters into a trembling, bitter laugh.
(stumbling backward, voice cracking) (trailing into a jagged sob that she swallows hard)
The performance across both moments should feel like a solitary woman's intimate stillness shattered into a feral, wounded animal response.
What to say (goes in MOSS text field): This lock's been dead for years... the hinges are the only things that remember how to sing. Oh, you think that's funny, do you? Just... just a little janitor?
Full DramaBox prompt (how to say — with the "quoted" dialogue):
A 62-year-old man with a gravelly, weathered baritone, delivering this high-quality studio voice recording with no background noise.
The same voice is hunched forward, fists up, spit flying as he shoves past a trash can toward a younger attacker, contempt raw and animal in his throat.
(spitting between words) "¡Métete con alguien de tu tamaño, pedazo de mierda!" (swinging, breath hissing) "Te voy a partir la madre, escúchame bien."
CUT TO:
The same voice now crumpled on one knee, ribs on fire, jaw slack with pain as he tastes blood, exhaustion cracking his snarl into something colder, quieter.
(coughing, grabbing his side) "Te... te lo juro." (voice dropping, venom trembling) "Si le tocas un pelo más... te mato."
The performance across both moments should feel like a single brutal breath—venomous fury collapsing into a wounded, whispered threat, every grunt and pause earned by real damage.
Direction only (dialogue removed — this is what prompt-sim compares against):
A 62-year-old man with a gravelly, weathered baritone, delivering this high-quality studio voice recording with no background noise.
The same voice is hunched forward, fists up, spit flying as he shoves past a trash can toward a younger attacker, contempt raw and animal in his throat.
(spitting between words) (swinging, breath hissing)
CUT TO:
The same voice now crumpled on one knee, ribs on fire, jaw slack with pain as he tastes blood, exhaustion cracking his snarl into something colder, quieter.
(coughing, grabbing his side) (voice dropping, venom trembling)
The performance across both moments should feel like a single brutal breath—venomous fury collapsing into a wounded, whispered threat, every grunt and pause earned by real damage.
What to say (goes in MOSS text field): ¡Métete con alguien de tu tamaño, pedazo de mierda! Te voy a partir la madre, escúchame bien. Te... te lo juro. Si le tocas un pelo más... te mato.
Full DramaBox prompt (how to say — with the "quoted" dialogue):
A child with a breathy, strained voice, delivering this high-quality studio voice recording with no background noise.
The same voice is flushed and dizzy from being bent over, focused entirely on a stubborn knot in their shoelace.
(breathing hard, words squeezed out) "No te muevas, tonto." (fingers fumbling) "Casi, casi está saliendo."
CUT TO:
The same voice now hot with frustration, a whine creeping in as the knot tightens and their arms start to shake.
(a sharp exhale) "¡Ay, ya! ¿Por qué no te desatas?" (voice cracking with complaint) "Me duele la cabeza y esto no se mueve nada."
The performance across both moments should feel like a real child's spiraling mood, from concentrated muttering to helpless whining, all while physically stuck in an uncomfortable pose.
Direction only (dialogue removed — this is what prompt-sim compares against):
A child with a breathy, strained voice, delivering this high-quality studio voice recording with no background noise.
The same voice is flushed and dizzy from being bent over, focused entirely on a stubborn knot in their shoelace.
(breathing hard, words squeezed out) (fingers fumbling)
CUT TO:
The same voice now hot with frustration, a whine creeping in as the knot tightens and their arms start to shake.
(a sharp exhale) (voice cracking with complaint)
The performance across both moments should feel like a real child's spiraling mood, from concentrated muttering to helpless whining, all while physically stuck in an uncomfortable pose.
What to say (goes in MOSS text field): No te muevas, tonto. Casi, casi está saliendo. ¡Ay, ya! ¿Por qué no te desatas? Me duele la cabeza y esto no se mueve nada.