/voice-realism

Voice Realism

Image quality in AI video is nearly past the uncanny valley — speech is not. Voice is the #1 tell that a video is generated. This skill rewrites prompts so the voice track survives scrutiny, and prescribes the post-processing that removes what prompting can't.

The core mindset: write the prompt like literary prose with epithets, not a dry algorithm. Generic elements produce generic (robotic) speech; details and shading produce a performance. A voice model infers pacing, pitch, and micro-emotion from everything around the line — so the job is to surround every line with context.

Workflow

Given an input prompt (or script), do these steps in order:

1. Parse the input

Identify:

• Target model — Veo, Seedance, Kling, ElevenLabs, or unspecified. Syntax rules differ (step 3). If unspecified, ask or output a model-agnostic version plus per-model dialogue syntax notes.
• Every spoken line — dialogue, voiceover, narration.
• Scene context — who is speaking, where, what just happened, what they feel. This drives the automatic parameter choices in step 2.
• Clip duration, if set — needed for the pacing check in step 4.

2. Choose the delivery automatically (don't ask the user)

Derive emotion, volume, and vocal texture from context. The user gives the scene; the skill picks the performance. Defaults by mood:

The whisper cheat. Synthesis artifacts are nearly inaudible in whispered speech — no pitch contour to get wrong, no resonance to ring. If the scene plausibly allows it (intimacy, secrecy, ASMR, night), prefer moving the line to a whisper. If a generated voice keeps "reading as AI" after other fixes, rewriting the scene so whispering is natural is a legitimate move.

Vocal texture is also free realism: an accent, a slight rasp, a smoker's voice, an age — one clause in the prompt, big payoff. Add one when the character description supports it.

3. Rewrite each line — the bundle rule

Never write bare dialogue. Every line gets rewritten as emotion + physical action + line, in one breath:

❌ She says: "hi" ✅ She smiles wearily, looks away, and quietly says: "hi"

The model adjusts facial animation, tempo, and intonation to match the description. Tone words (weary voice, shouts excitedly, barely audible) are the cheapest way to kill robovoice.

Model syntax — get the dialogue delimiter right or the model may read stage directions aloud:

• Veo: line comes after a colon — she says: hello there
• Seedance: line in double quotes — she says "hello there"
• ElevenLabs v3: supports inline audio tags — [whispers], [sighs], [laughs], [clears throat], and custom tags (they're interpreted semantically, so descriptive custom tags like [tired chuckle] work). Place them where the sound happens: Another long day [sighs]. There's also an auto-tag mode where the model sprinkles tags over plain text for you — fine as a first pass, then hand-tune.
• Other / newer models: check current docs before assuming — delimiter conventions change between versions.

Write like people talk, not like people write:

• No long, complex sentences — no nested clauses
• Filler words, hesitations, "uh", "hm" where natural
• Audible breaths in and out
• Skip heavy punctuation, but ellipses and dashes are excellent — they read as pauses and trail-offs

4. Pacing check

If the clip duration is fixed, count the words. Natural speech runs roughly 2–2.5 words/second English (casual is slower). A 5-second clip fits ~10–12 words; more than that and the model compresses delivery into a 4× voicemail. Cut the line or extend the clip — say which you did.

5. Bake acoustics into the prompt

The model has heard rooms during training — describe the space and it will render the reverb:

• wind audible in the microphone, small 10 m² room, speaking loudly in a vast vaulted cathedral, 20 m ceilings
• Match acoustics to the visual scene — a street shot with studio-dry voice is an instant tell.

Seedance 2.0 generates video + audio in one pass and accepts an explicit mix priority — always include one when targeting it, e.g.: dialogue clean and prominent, music low, ambient subtle Without it, layers turn to mush.

6. Frame for lipsync

If the character speaks on camera, constrain the shot in the prompt:

• Medium or close-up shot — on wide shots lips don't get enough resolution and start living their own life
• Static camera while the line is delivered
• Face frontal or three-quarter to the camera

If the source prompt asks for a wide shot with dialogue, either tighten the shot or move the line to voiceover — flag the change.

7. Output

Return exactly this structure:

## Rewritten prompt
[the full rewritten prompt, ready to paste]

## What changed and why
[3-6 bullets: delivery chosen and the context cue that drove it,
syntax fixes, pacing math, acoustics added, framing changes]

## Post-processing checklist
[the subset of the post chain below that applies to THIS video,
with the reasoning — e.g. which background noise matches the scene]

Post-processing chain

Prompting gets ~80% there; the last 20% is de-idealizing the audio in post. Recommend from this menu (tailor to the scene, don't dump the whole list blindly):

1. Cut 2–4 kHz a few dB — this band carries the characteristic "AI ring" / synthetic sheen.
2. Room reverb, 5–10% wet — restores the acoustic space that synthesis flattens. Match the room size to the visual.
3. Background noise bed at ≈ −50 dB — pick the bed from the scene (street rumble, room tone, café murmur, wind). A voice with zero floor noise sounds like a studio, and nothing in real life is a studio.
4. Voice morpher at very low intensity — "phone call" or "bad microphone" style degradation. Barely perceptible settings only; the goal is imperfection, not an effect.

The unifying principle: remove polish. Real recordings are slightly wrong in a dozen tiny ways — pick 2–3 of those ways and add them back.

Worked example

Input (target: Seedance, 5 s):

A woman comes home and says: I got the job. Wide shot of the apartment.

Rewritten:

Medium close-up, static camera, her face in three-quarter view. A woman in her 30s pushes the door shut behind her, drops her keys, exhausted but glowing. She exhales, breaks into a shaky laugh, and half-whispers "I got... I got the job." Small apartment hallway, ~8 m², soft evening light, her voice slightly muffled by the close walls. Audio: dialogue clean and prominent, ambient room tone subtle, no music.

What changed: wide shot → medium close-up + static camera (lipsync); bare line → emotion + action + line bundle; delivery derived from context (big news + coming home tired → shaky laugh + half-whisper, which also masks synthesis artifacts); Seedance double-quote syntax; 7 words ≈ 3 s of speech — fits 5 s with the physical business; room size stated for acoustics; mix priority appended.

Post for this clip: −3 dB dip at 2–4 kHz; small-room reverb ~7% wet; apartment room tone + faint street bleed at −50 dB.

Scope notes

• The job is always to make the voice better, never to remove it — don't suggest going voiceless; the user chose a speaking concept deliberately.
• This skill rewrites prompts and prescribes post steps; it doesn't run generation. Pair with gen_video.py / render_video.py from tools/ when the user wants the video actually produced.