Native dialogue is the feature that finally made AI video feel like filmmaking instead of silent-film generation with a voiceover glued on in the edit. It’s also the feature most likely to waste your credits in a new and exciting way.
A line that reads perfectly on the page comes back mumbled, misattributed to the wrong face, delivered in the wrong emotion, or cut off mid-word because the clip ended. None of that is bad luck. Almost all of it is formatting — how you quote the line, where you place it, and how much you ask one generation to say.
In my pipeline, dialogue is boarded like camera work — the habit at the heart of the storyboard-first workflow. One speaker per panel, one beat per panel, the line written on the board where the mouth will be. That discipline came from live-action habit, and native-audio models reward it exactly the way a sound recordist would.
This guide covers the writing side: quoting speech for models that generate voice with video — principally Veo and Kling in their current native-audio modes — and knowing when to write the line and generate silent anyway.
The Model Reads Your Punctuation As Direction
With silent generation, punctuation was style. With native dialogue, punctuation is stage direction the model can hear. Quotation marks define what’s spoken versus what’s described. A colon before a quote reads as an introduction. Ellipses become pauses. Exclamation marks raise volume and energy whether you wanted that or not.
I’ve had a character shout a line I meant as weary, purely because my draft ended it with an exclamation mark out of typing habit. The model, quite reasonably, performed the punctuation I gave it. Yeah. It does that.
Treat the quoted line as the literal script the actor receives. Read it aloud, once, at natural speed, before you generate. If you stumble, the model will too — it has no rehearsal, no director’s note, no second take inside the same generation. Tongue-twisters, stacked consonants, and long subordinate clauses that work in prose trip synthetic delivery the same way they trip a first table read.
Capitalization inside quotes also carries weight. ALL CAPS reads as shouting. A lowercase, unpunctuated line often comes back flat and mumbled, as if the model inferred low energy from low effort. Standard sentence case, honest punctuation, and line breaks between speakers are not pedantry here. They’re the cheapest performance notes you can give.
One more formatting truth that saves repeat pain: keep descriptive text and spoken text cleanly separated. “She says sadly, ‘I knew it'” buries the performance note inside narration where models weight it inconsistently. “She pauses, shoulders dropping. Then, quietly: ‘I knew it.'” gives the model a visible beat, then a line with its delivery attached to the quote itself.
Structure the sentence the way you’d structure the moment. The model stages sentences more literally than most writers expect.
Quoting Lines: Placement And Length Limits
Placement first. Put the spoken line after the subject and the immediate action are established, and keep it in the same sentence block as its speaker. A reliable shape: subject and beat, speaker tag, colon, quoted line. For example: “Maya sets the envelope on the counter and looks up. Maya, steady: ‘You knew about this.'”
The model meets the person, sees the beat, then hears the words attached to the right face. Lines placed at the very start of a long prompt, before the speaker is described, are the ones that wander onto the wrong character in two-person frames. Seen it happen. Multiple times.
Length is where credits go to die, so let me be blunt about the arithmetic. Spoken English runs at roughly two to two-and-a-half words per second in natural delivery. A five-second generation can hold about ten honest words, with room for a breath and a reaction. An eight-second generation can hold perhaps eighteen, if nothing else is competing.
Writers routinely paste forty-word lines into eight-second clips and then blame diction. The model didn’t mumble your monologue. It was forced to either race it, truncate it, or smear it — and it chose, as models do, without asking you. Right?
Write lines at the length the face can perform. Short lines also act better. “I’m not going.” lands. “I’ve been thinking about everything that happened and I don’t think I can go through with it anymore” is a voiceover sentence wearing a dialogue costume. If the thought is long, split it across shots and let reaction frames carry the middle — which is how dialogue is edited in real films anyway. The close-up says the first half, the listener’s face carries the pause, the close-up finishes. Your storyboard already knows how to do this. Trust it over the single heroic take.
Practical length rules from my own shot lists: one line per generation, one sentence per line where possible, and if a line needs a comma’s worth of pause in the middle, consider whether that comma is actually a cut. It usually is.
Naming The Speaker Without A Cast List
Native-dialogue models don’t receive your character bible. They receive this prompt, this frame, and whatever naming you give them inside it. Speaker attribution has to be rebuilt, briefly, in every prompt that contains speech.
The cleanest method is a consistent speaker label tied to a visible identifier. Not just a name — a name anchored to what the camera sees. “The woman in the red coat (Maya): ‘…'” outperforms “Maya: ‘…'” in any frame where Maya hasn’t been introduced in that same prompt, because the model can bind the label to a figure.
In text-to-video, introduce the speaker in the descriptive sentence first, then tag the line with the same words. In image-to-video, the frame carries the identity, but the tag still tells the model which of the visible people opens their mouth — an instruction that matters enormously in two-shots.
Consistency across a sequence matters as much as clarity within one prompt. If she’s “the woman in the red coat” in shot 1 and “the brunette” in shot 4 and “Maya” in shot 7, the voice may drift along with the label — and the face is being re-negotiated each time too. My shot lists carry the exact speaker phrase, copied, not retyped from memory. It feels fussy. It prevents the slow identity slide that the character consistency workflow exists to fight, now extended from faces to voices.
Avoid pronouns as the only speaker tag. “She says: ‘…'” in a frame with two women is a coin toss conducted with your credits. Repeat the noun. It reads slightly stiff on the page and stages perfectly in the generation — and staging is the point.
Also: don’t name speakers who aren’t speaking. Listing three characters and quoting one invites the model to audition the others. Cast the speaking shot tightly, the way the board does — the speaker, and at most a listener whose silence is part of the beat.
Parentheticals: Tone, Pace And Pauses
Parentheticals — the small delivery notes attached to a line — are the highest-leverage words in a dialogue prompt, precisely because there are so few of them. Used well, they give you performance control. Used carelessly, they contradict the words and the model averages the contradiction into mush.
The parentheticals that reliably change delivery are the physical and temporal ones: (quietly), (quickly, before he can answer), (a beat, then), (laughing despite herself), (calling from the doorway). Each describes something a body or a clock does.
The ones that wobble are abstract emotional labels stacked on already emotional lines: “(devastated, heartbroken, utterly destroyed): ‘You lied to me.'” Three synonyms for the same feeling don’t triple the performance. They inflate it toward melodrama. Same failure as in expression writing generally — more on that elsewhere.
One parenthetical per line is my hard rule, two only when one is timing and one is volume. (Whispered, after a pause) earns both slots. (Angrily, sadly) spends two slots contradicting each other. See what I mean?
Pauses deserve their own mention because they’re the note creators most want and most often mistype. An ellipsis inside the line — “I thought… I thought you’d left” — produces a genuine hesitation when the line is short. In a long line it produces drift, because the model stretches uncertainly.
A cleaner tool for a real pause is staging it before the line: describe the beat, then give a shorter line. “He looks at the photograph. A long moment. Daniel: ‘Where did you get this?'” The pause becomes visible, the line stays performable, and the edit inherits a frame it can hold on.
Delivery notes also interact with casting. Asking for a whispered line from a character framed in a wide shot invites the model to either boom the whisper or shrink the performance into inaudibility. Match volume to framing the way a boom operator would. Whispers belong in close framing. Shouts belong where the body has room. The prompt can’t repeal acoustics, and pretending otherwise is another quiet way to burn a generation.
Two Speakers, One Clip: The Honest Limits
Every creator tries it: one clip, two characters, a snappy exchange. It’s the demo-reel promise of native dialogue. Here’s the honest production report from doing it repeatedly in both Veo and Kling: a single short exchange can work — one line each, clearly tagged, minimal movement — and everything beyond that degrades quickly and expensively.
The failure modes are consistent. Lines swap speakers. The second line starts before the first finishes, stepping on it. One character mouths both lines while the other becomes an audience member with a moving jaw. Emotional tone bleeds across the cut that never happened, so the reply carries the first speaker’s feeling instead of its own. Each added line multiplies these chances, because the model is staging, performing, and sound-mixing a scene it can’t rehearse.
If a two-speaker single clip is genuinely the right shot — a locked-off two-shot with a quick exchange, say — stack the odds. Keep both lines very short. Tag each speaker with a visible label, repeated, not pronouned. Put the speakers on clear screen sides and keep them there in the description: left and right are stage directions models respect more often than names alone. Remove competing action. A walking, gesturing, prop-handling conversation in one generation is three budgets spent on one clip.
And then, honestly, consider not doing it. The alternative is older than AI video and better: generate each speaker separately, each in their own framing, each with one performable line, and build the conversation in the edit. The scene feels more filmed, the dialogue is controllable line by line in Premiere, a bad line costs one regeneration instead of a whole exchange, and you gain reaction shots — the ingredient that makes generated dialogue feel directed.
My own dialogue scenes are assembled this way by default. Single-clip conversation is the exception I allow for throwaway exchanges, never for a beat the story depends on.
That approach also future-proofs the scene. Voices and models improve at different rates; a scene built from single-speaker clips lets you regenerate one performance later without restaging the whole conversation.
When To Write The Line And Generate It Silent Anyway
The least obvious skill in dialogue prompting is deciding not to use the dialogue the model makes. Writing the line into the prompt and generating with native audio has real value even when you intend to replace the voice: the face performs the words. Mouth shapes, breath timing, emphasis, the small head movements that accompany speech — all of it is driven by the quoted line.
Strip the generated audio in the edit, lay in your chosen voice, and the sync relationship is close enough to sell, because the performance was genuinely organised around those syllables. I’ll be straight with you — this trick has saved me more than once.
Generate-silent-anyway is the right call in several recurring situations. When the voice matters more than the convenience — a series character whose voice is cast, a client’s own voice, any line that must match previously published episodes. When the line is longer than the clip can honestly hold as speech, but you still want the face to perform its rhythm while a treated voiceover carries the words. When the accent, language, or delivery is specific beyond what you’re willing to gamble a generation on. And when the scene’s sound design will be built in post anyway, generated location audio and all.
The converse discipline matters too: don’t write dialogue into prompts for shots with no speaker. A quoted line in an establishing shot or a product shot gives the model an instruction with no mouth to attach to, and it will sometimes invent one. Silent shots should be textually silent. The board should say so, and the prompt should honour it.
My working default, offered plainly: write every line on the board, quote it in the prompt when a face performs it, judge the generated voice with genuine openness — native voices keep surprising me on short, well-formed lines — and keep the edit free to replace anything that doesn’t serve the scene.
The line is part of the shot’s direction. Whether it’s part of the final audio is a separate, later, cheaper decision. Directors have always shot lines they later looped. Generating them is no different, and it’s a much better use of the feature than either avoiding dialogue entirely or betting scenes on a single synthetic take.