AI video dialogue prompts: give each line a speaker.
Write who speaks, the exact line you want, and how the other person responds. Keep speech instructions separate from background sound and visual action. Then listen to the exported clip and check who actually says each word: a good-looking talking face does not establish that the dialogue is correct.
1. Decide where the spoken performance comes from
The right starting point depends on whether you are still inventing the performance or already have approved speech. Check the selected tool and mode before writing; a video generator does not automatically accept audio input or generate dialogue.
- Generate speech with the picture: describe the scene, identify the speaker and supply the line in a mode that supports spoken audio. Review the resulting words and performance together.
- Drive a face from approved audio: choose a workflow that accepts your voice track and a compatible image or clip. Check the actual input requirements before preparing a long recording.
- Use voiceover: place the approved track over suitable footage in an editor when the speaker need not appear on camera. Mouth synchronization may be unnecessary for that shot.
Runway's current lip-sync overview distinguishes prompt-generated talking characters, audio-driven images, and performance transfer. These routes use different inputs. If you need to preserve a particular recorded delivery, check an audio- or performance-driven route rather than expecting a text description to reproduce it exactly.
Use faces and recordings you have permission to use. For the exercises below, the people and lines are fictional. No reference image, voice recording or generated result has been supplied or tested.
2. Save the line before expanding the scene
DeepMind's Veo 3 prompt guide distinguishes giving characters a discussion topic from giving them specific words. Choose which result you need. “They discuss the plants” leaves the wording open; an approved script needs its own written lines and an output check.
Speaker identifier and visible position; approved line; language; intended delivery; who listens; pause or handoff; relevant action; background sound; caption text for the edit; audio source if one is being supplied.
Read the line aloud at the intended pace. Include the listener's reaction and the pause when judging whether it fits the selected clip duration. There is no universal word count that guarantees natural delivery. Shorten the exchange or plan separate shots when it feels rushed, rather than asking the model to squeeze in more speech.
Google Cloud's current video prompt guide recommends separate sentences for audio directions and distinguishes dialogue, ambient noise and individual sound effects. Keep those jobs clear in your brief. Adapt the punctuation and any dedicated dialogue fields to the chosen service; this card is a writing aid, not a universal API format.
3. Start with one speaker before adding a reply
These original prompts are proposals for a speech-capable generation mode. They are not measured examples or a promise of exact wording. Keep the approved script alongside each actual take.
Example A: a single speaking character
A gardener in a blue apron stands beside a seedling bench in a bright greenhouse. The camera holds an eye-level medium view of the gardener. The gardener looks toward a visitor just outside the frame and says in English, with a calm, pleased delivery: “The seedlings are ready.” The gardener then pauses. Quiet greenhouse room tone sits underneath the speech.
Review whether the line finishes clearly and whether the mouth and face remain usable. In an audio-driven workflow, put the approved line into the recording and follow that tool's visual-prompt instructions. Typing a line into an unrelated field does not supply an audio track.
Example B: an ordered two-person exchange
An eye-level medium two-shot at a greenhouse seedling bench. The gardener in the blue apron stands on the left. The visitor in the tan coat stands on the right. First, the gardener says in English, calmly: “The seedlings are ready.” The visitor listens. After the gardener finishes, the visitor replies in English, warmly: “I'll bring the trays.” The gardener listens during the reply. Quiet greenhouse room tone continues beneath the exchange. The camera holds its view.
The identifiers connect each line to a visible person, and the handoff makes the intended order explicit. This still needs verification. If bodies, props or positions become confused, use the two-character action guide to simplify the visual plan before adding more dialogue directions.
4. Review the words before judging the polish
- Listen without reading the script. Write down what you actually hear, then compare it with the approved line. Check for missing words, invented speech, wrong language or a clipped ending.
- Watch who speaks. Confirm that each line belongs to the intended person and the listener behaves as planned. A correct sentence from the wrong face is still a failed take.
- Check the mouth at normal speed. Inspect visible starts, stops and pauses, then look more closely at any suspect moment. Mouth motion and audible speech should read as one performance.
- Check the sound bed and final edit. Listen for ambience or music masking the words. Add and proofread intentional captions in the edit; generated on-screen text does not prove that the soundtrack matches the script.
Save the source filename, exact prompt or audio input, model and mode, settings, observed error and decision. Compare the exported file, not just an attractive preview. For voiceover-led footage, the B-roll planning guide helps connect the pictures to the narration.
5. Change the part that actually failed
- The line is wrong: shorten or clarify it, or test an approved-audio workflow if your selected service supports one.
- The wrong person speaks: clarify speaker identifiers and turn-taking. If that remains unreliable, plan one speaker per shot and assemble the exchange using a storyboard.
- The words are right but synchronization fails: in an audio-driven route, inspect the supplied recording and visible face. Runway's overview recommends clean audio and a clearly visible face. Reject the take if the performance still does not work.
- The dialogue is buried: simplify background sound or adjust the mix in an editor when the available tracks permit it.
Organize the scene instructions in Shotmelo's manual prompt planner, then save a local draft or export its text. Keep the approved script with the plan. The current preview does not synthesize voices, render video or perform lip sync. Use the selected external service and an editor to produce and review the actual dialogue.