# Make a talking-character launch video with AI Turn a feature announcement into a 21-second video where two characters talk it through, lip-synced to one dialogue take. Includes the script, cast and shot prompts. Author: John Holliman Canonical: https://voyager.so/blog/talking-character-launch-video Published: 2026-10-16T23:45:00-07:00 Updated: 2026-10-16T23:45:00-07:00 A feature announcement usually arrives as a screenshot and a paragraph. Most people scroll past both. Give the news to two characters instead. They trade short lines, each line makes one claim, and the feed shows something that moves and talks. For this example, two dolphins announce a voice integration in about 21 seconds. ![Two dolphins in a glossy aqua world trade lines about new voices, with captions, feature cards and an end card.](https://cdn.moda.app/voyager/blog/talking-character-launch-video/hero-a46a2856c0b43b82.webp) [Video: Two dolphins in a glossy aqua world trade lines about new voices, with captions, feature cards and an end card.](https://cdn.moda.app/voyager/blog/talking-character-launch-video/film-dc994220449555b9.mp4) Seven lip-synced shots over one dialogue take, with captions and feature cards so it also works muted. Sound on for the voices. Narrator dolphin: Welcome to Voyager voices! Second dolphin: Fish Audio voices? In Voyager? Narrator: Yep! Including featured voices, licensed from real creators. Second dolphin: With my own Fish Audio key? Narrator: Bring your own key! Settings, Accounts. Second dolphin, whispering: Wait, is this whole scene one take? Narrator: Fish Audio Drama 3. Every character, one take. End card: Voyager, now with Fish Audio voices. **The whole video hangs on one recorded conversation.** Generate the dialogue as a single multi-voice take first, then make one shot per line and sync each mouth to its slice of that take. The pacing, energy and handoffs between the characters come from the voice track, not from seven separately generated clips. ## What you need * **Skills:** `make-ai-video`, `voiceover`, `edit-in-remotion` and `make-music`. * **Plan:** Creator or higher, because the shots use Seedance 2.5. Voices and music are included from Starter. * **Tools:** Remotion for the captions and cards, and FFmpeg to conform the shots. * **Made in Voyager:** this example is a real Voyager run of the workflow, from the script to the final mix. ## Script to the length you want Write the lines before anything else, and count them against the clock. Our first script ran about 30 seconds; the target was 20. A raw voice take also runs 25 to 50 percent longer than the final edit, so start short. Give every line one claim and say it plainly. "Featured creators" and "bring your own key" confused people until the lines spelled them out: voices licensed from real creators, and a key you add in Settings. End on the brand line, and put a small twist before it. Here, a whisper asks whether the whole scene is one take. ```text Make a fun ~20-second 4:3 announcement video for [brand]'s [product] for [platform]. Two characters, [character 1] and [character 2] ([look]), talk through it: [feature 1], [feature 2], [feature 3]. Style: [style]; use my references for palette only. It should feel like a video, not a slideshow, with on-screen text and motion graphics. Don't mention [names to avoid]. Budget: [limit]. Show me the lines before anything is generated. ``` Approve the script before you spend anything. Every later step depends on it. ## Record one dialogue take Ask for the whole conversation as one multi-voice request, with a direction tag on each line: bright and quick, whispered, a soft laugh. Make two takes, transcribe both and keep the one that reads every line correctly. Then tighten the gaps between lines, but leave the whisper's own pauses alone. ```text Voice the approved script as one dialogue take. Use [voice 1] for [character 1] and [voice 2] for [character 2], switching at each line. Add a delivery direction to every line. Make two takes, transcribe them, and keep the one where every line reads correctly. Aim for 20.5 to 21.5 seconds after tightening the gaps. ``` ## Cast the characters once Design each character as a Voyager Cast card: a hero portrait and a full-body view made from the approved hero, with a fixed written description. Every shot then receives the same cards, so the characters stay recognizable from line to line. ![Two Cast references side by side: a lavender-silver dolphin with a glass headset microphone and a teal dolphin splashing in bright water.](https://cdn.moda.app/voyager/blog/talking-character-launch-video/cast-f951f12c743e5ebf.webp) The two Cast heroes. Each card also holds a full-body view and a fixed description that every shot reuses. Our first characters read as too childish for the announcement. The fix was in the description: realistic anatomy, not a cartoon, not cute. Non-human faces looked uncanny head-on, so every shot uses a side or three-quarter profile with the eyes off the lens. ## Make one shot per line Generate one background plate per scene from text, leaving the right third clear for type. Use any style references for palette and lighting only; plates that copied a reference's composition were flagged. Each shot then gets its plate, both Cast cards and that line's audio slice. We used Seedance 2.5 reference-to-video at 720p and 4:3, four or five seconds per line, and wrote the prompt in timed beats: ```text @Image1 is the setting; do not copy any characters from it. [Character] is in [action], filmed in three-quarter profile with [camera move]. 0-1 seconds: [framing]. The line starts within the first second. 1-4 seconds: [character] says in English, [tone], at the exact pace of @Audio1: "[line]". @Audio1 is the speaking character's dialogue; sync the mouth to it with small, natural jaw movement. Dialogue and light ambience only: no music, subtitles, text, logos or watermark. Keep each character consistent with its reference. ``` Check the estimate before submitting. At the time, a 720p Seedance shot of about five seconds was quoted at roughly $2 at list price, and each took about five minutes; the shots ran in parallel. A content filter blocked a few harmless shots, including the whisper, without a reason. A gently reworded retry passed. ## Conform the shots to the voice track Measure where speech starts in each clip and compare it with its slice. Trim, and speed-match only slightly; we capped it at about 1.25 times. A clip that needed far more was reshot with "talks from the very first second". Then cut every shot to an exact frame count with FFmpeg so the shots butt together, and lay the original take over the whole edit. Two glitches only appeared at cuts. Rounding each shot separately left a blank frame, and a generated clip can open on a different angle for a few frames before its real shot begins. Scan each clip for internal cuts and start after them. ## Make it work with the sound off Feeds autoplay muted, so the text carries the story too. We built the graphics in Remotion: corner tags, an opening title, a feature card timed to each claim, a caption for each line, the reveal title and the end card. Keep overlays off faces, and finish each fade before its cut. Last, a light music bed that drops to a held pad under the whisper and hits on the reveal. The whisper was still buried in the first mix; raising it about 4 dB and dipping the music fixed it, which a transcription of the final mix confirmed. One claim needed rewording at the end. "One take" stays true of the voices, not of a picture made from seven shots. Check every line of the copy against what the viewer actually sees. Swap in your own characters and claims, and **let one conversation carry the whole video.** ## Make your next project in Voyager Create an account, download Voyager, and start making. [Sign up](/sign-up)