In shortSeedance, Kling and Veo render each clip with its own audio, and that audio includes music written for that clip alone. Stitch eight clips together and you get eight songs that stop and restart at every cut, in moods that do not connect. The fix is to remove the music from each clip while keeping the dialogue and effects (stem separation), then score the whole timeline once in your editor, change the music only where the story turns, and duck it under every line.
Seven Seedance clips cut into one scene: as rendered, each clip brought its own music and the sound jumped at every cut.
Why the music breaks at every cut
AI video models make short clips, usually 5 to 15 seconds and rarely more than 30. When a model generates sound with the picture, it scores each clip on its own: it hears the prompt, picks a mood and writes a few seconds of music that start with the clip and end with it. It has no idea what came before or what comes next.
That is fine for a single clip. It falls apart the moment you cut clips into a scene:
The music stops and restarts at every cut. Each clip's music begins from nothing, so the track drops out and comes back every few seconds.
The mood jumps. Clip 3 decided the scene was tense, clip 4 decided it was romantic, clip 5 went epic. A viewer hears three different films in twenty seconds.
Tempo and key do not match. Even when two clips pick the same mood, the beats land in different places and the keys clash across the cut.
The level jumps. Some clips mix the music loud, others quiet, so the volume pumps from shot to shot.
A cut that is invisible in the picture becomes obvious in the sound. On a phone, where people watch with sound on and swipe fast, that seam is often what makes an AI scene feel cheap.
As rendered, each clip carries its own music, so the track drops at every cut. After the fix, the clip music is gone, the dialogue stays, and one score runs across the cuts.
Can you avoid it when you generate?
Partly. Two things help, neither solves it:
Ask for no music. Adding "no background music, dialogue and natural sound only" to the prompt lowers how often a model adds a score. Models still add it often enough that you cannot count on it across a whole episode.
Generate without audio. Some tools let you switch sound off. You then lose the dialogue and lip-synced voices the model made, and have to rebuild every line. For drama, that usually costs more than it saves.
So for most projects the music has to come out after generation.
Step 1: strip the music from each clip, keep the dialogue
This is stem separation: splitting a mix into voice, music and everything else, then keeping only the parts you want. Do it clip by clip, before you cut, because each clip's music is different.
Open source: Demucs, and Ultimate Vocal Remover, which wraps Demucs and similar models in a desktop app. Both run locally and give you a vocals stem and a music stem.
In editors: DaVinci Resolve Studio has voice isolation, CapCut has vocal isolation, and Adobe Audition and Premiere Pro offer speech enhancement that pushes music down.
Three things to watch:
Only clean clips that actually have music. Running separation on a clip with dialogue and footsteps alone can thin the voice for no gain. Listen first.
Keep the non-music sound if you can. A pure "vocals" stem throws away the door, the glass and the rain the model made. If your tool gives you a separate "other" or effects stem, keep it with the voice.
Check for artifacts. Where the music was loud under a line, separation can leave the voice watery or metallic. If one clip is badly damaged, it is often cheaper to regenerate that one shot than to fix it by hand.
Step 2: score the whole timeline, not the clips
Now put the cleaned clips in order and treat the scene as one piece of music:
Find the story turns. Watch the cut and mark the moments the scene changes: a standoff becomes a fight, a fight becomes a rescue, an argument becomes a confession. Those are your music changes. Everything between two turns gets one track.
Keep sections long. A section shorter than about 8 seconds feels like a mistake rather than a change. If two turns are closer than that, keep the music running through the first one.
Same mood, same track. If the scene returns to the standoff, bring back the standoff music instead of picking a new one. Repetition is what makes it feel scored.
Run the music across the cuts. The point of the fix is that the music no longer cares where the picture cuts. Put music changes on story turns, not on edits.
Crossfade the changes. One to two seconds between sections, or cut on a strong hit when the turn is violent.
Use silence sparingly. A second or two of no music right after a shock, such as someone being struck down, can land harder than any cue. Once or twice per video is plenty.
Step 3: duck the music under the dialogue
The lines have to win. In most editors you can set this up once:
Duck by about 10 dB whenever someone speaks, and bring the music back up in the gaps. Premiere Pro (Essential Sound, Ducking) and DaVinci Resolve (Fairlight sidechain on a compressor) can do it automatically from the dialogue track.
Keep the ramps short, about a quarter of a second down and half a second back up, so the music does not swell in the middle of a sentence.
Mix for phones. Check the finished scene on a phone speaker at low volume. If you have to strain to hear a line, the music is still too loud.
Fill the holes the old music left. After stripping, some clips sound empty between lines. A quiet ambience bed per location (a room tone, rain, a crowd) glues the cuts together under the new score.
Long videos: do it in passes
Everything above scales to a full episode, but separation and scoring by hand get slow past ten minutes. Work scene by scene, keep a short note of which track belongs to which mood so the same mood keeps the same music across scenes, and do the final level check on the whole episode, not per scene.
The one-click option
Dawnwood's Auto Sound does all three steps for you. Upload the stitched video on the Auto Sound card in Tools, or press Auto sound in the editor on a timeline of Seedance, Kling or Veo clips. It strips the music from each clip only where music is actually heard, keeps the dialogue and the sound already there, then watches the whole timeline and lays one continuous score that changes only where the story turns, with sound effects and an ambience bed added where the picture needs them. The music ducks under every line, and every effect lands on the editor's Sound track so you can swap, nudge or reject it.
It handles long videos too: each pass covers up to 15 minutes of new footage, and you can add the next part of an episode and run it again, paying only for footage it has not heard before.
Once the music holds across the cuts, the last step is getting the video out. Finish & post adds captions and a cover, drafts the post text, and posts the video to your own TikTok, Instagram or Facebook without leaving the page, at no extra cost. Your first 3 videos post publicly without a card; for YouTube, download the packaged video and upload it yourself. See the Auto Sound page for a before and after on seven stitched clips.
Free course · Level 2 · Build a season · Lesson 13 of 23
Fix the music in your stitched clips in one upload
Dawnwood's Auto Sound takes a video stitched from AI clips, strips the music baked into each clip while keeping the dialogue, lays one continuous score across the whole timeline that ducks under every line, and adds the missing effects and an ambience bed per scene. Upload a video and get it back with sound in one pass; a failed run is refunded.