Captions, voice and music for AI video clips: the order to do it in

Most guides on this topic assume you are starting from a long recording. AI clips are different. A clip from a video model such as Seedance (ByteDance's video model) or Kling (Kuaishou's video model) usually runs a few seconds, up to roughly 30 seconds on the longest settings. You will join several of them. If you edit the clips first and add sound afterwards, you end up regenerating clips so they fit the words. This page gives the order that avoids that, plus the checks that keep you out of trouble with music rights and platform rules. I use Higgsfield, a web service that puts many image, video and audio models in one place, as the example where it helps. A free editor does some jobs better.

This page may contain affiliate links from which Blur Studio earns commissions.

The short answer: build the audio first, then cut clips to it

The order I recommend:

1. Lock the script. Read it aloud and time it.
2. Make the voice from that script.
3. Lay the music under the voice and set the levels.
4. Export that finished audio bed at full length.
5. Cut or generate clips to match the bed, not the other way round.
6. Add captions from the approved script.
7. Preview on the real platform, sound on and sound off.

The reason is simple. A voice line has a fixed length. A clip can be trimmed, sped up slightly or covered by the next shot. If the audio is final, every clip you generate has a clear job: cover these four seconds. You also stop paying for clips you then cut away. If you want the cost side of that, see how to spend Higgsfield credits wisely.

What you are assembling

Layer	What it is	Who owns the risk
Clips	The AI-made or filmed pictures	You, for rights and accuracy
Voice	Narration, a lip-synced character or both	You, for consent and disclosure
Music and effects	A bed under the voice	You, for licence terms
Captions	Words on screen or a caption file	You, for accuracy and placement

Each layer fails in its own way, so the rest of this page takes them in the order you should build them.

Step 1: write the script and lock the voice

Write for the ear. Short sentences. One idea each. Read the script out loud once and mark where you breathe. A rough guide many editors use is about 150 spoken words per minute, but treat that as a habit, not a rule. Your own read-aloud time is the real number.

For the voice you have three options:

• Record yourself. A phone in a quiet room, held close, beats a poor synthetic voice. It also removes any question about whose voice it is.
• Text to speech. A synthetic voice reads your script. ElevenLabs (a voice AI service) is the best known. See the ElevenLabs review, and for choosing and tuning a voice see how to choose an AI voice for company videos.
• Voice change. You record a rough read in your own voice and a tool swaps it for another voice, keeping your timing and emphasis.

Cloning a real person's voice needs their written agreement. If you want that route, read the voice cloning page and the consent checklist first.

If your whole video is mostly narration over footage, an editor that puts voice, captions and music on one timeline is simpler than juggling files. This guide to the ElevenLabs Studio route covers that, and for most narrated videos it is the easier path.

Step 2: decide what sound each clip keeps

Some video models make their own sound. Higgsfield's help center says Cinema Studio 4.0, which runs on Seedance 2.5, has native audio in all versions, and that the sound toggle on Seedance does not change the cost. That is useful when you want ambient sound or a character speaking in the clip. It is a problem when you plan to overlay your own voice, because you will hear two sources.

Situation	Keep the model's audio?	What to do
You will overlay narration	No	Switch the sound off if the tool allows it, or mute the clip in the editor
A character speaks on camera	Maybe	Check the words are exactly your script; generated speech often drifts from it
You want a lip-synced character with your voice	No	Mute generated speech, then add your audio through a lip sync step
Ambient sound only (wind, street, room)	Often yes	Keep it low under the voice, or replace it with a library effect
Product shot with no speech	No	Use a clean music bed and one or two licensed effects

Lip sync means a tool changes the mouth movement on a face so it matches an audio track. Higgsfield's Lipsync Studio has an audio tab with text to speech, voice change, translation into 18 languages and a custom voice made from an MP3 or WAV sample, per its help center as of November 2026. The details are in the Lipsync Studio guide. Check the plan and the current page before you rely on any one model.

Step 3: add music you are allowed to use

Music is where small businesses get caught, because a track that is fine in a personal post can be refused or muted once it is in an ad. Pick the source before you pick the song. If the song is the whole video, see what is realistic for an AI music video.

Source	Good for	Watch out for
YouTube Audio Library (free tracks from YouTube)	Videos you post on YouTube	YouTube says only Audio Library tracks are known to it to be copyright-safe on its service, and some tracks need a credit in the description. That says nothing about other platforms or ads, so read the licence on each track.
TikTok Commercial Music Library	Business accounts on TikTok	TikTok's ads help page (read 4 November 2026) says businesses cannot use the general music library for commercial use and should use the Commercial Music Library for all commercial activity, and calls its songs free for businesses to use. It does not say what a track may be used for outside TikTok, so keep the library for TikTok posts and read the page again before you reuse a video elsewhere.
Meta Sound Collection	Posts inside Instagram and Facebook	Secondary sources say tracks are licensed only for use inside Meta apps, which would matter if you export the video for other places. Meta's own page is not confirmed here, so check it, especially for ads.
Generated music	Custom beds with no famous song attached	Plan terms decide what you may do with it. See ElevenLabs sound effects and music for how one service handles it.
A paid stock-music licence	Ads and client work	Check that the licence covers paid social ads and the countries you target.
"Royalty-free" labelled tracks	Sometimes anything	The label means you do not pay per play. It does not mean the track is licensed for ads or safe from a copyright claim.
A chart song	Nothing in a business post	No licence. Expect a mute, a claim or a removed ad.

For more on sources, see where to get background music without licence fees. Keep a note of where each track came from, the licence page and the date you checked it. If a platform asks later, that note is your proof.

Keep the music under the voice, and check the level on phone speakers.

Step 4: cut the clips to the audio

Now work from the finished bed. Mark each voice line on the timeline and give it a clip. Then:

• Cut on a word or a beat, not mid-word.
• Keep each clip close to the line it illustrates. A reader hears "waterproof" and should see water within a second or so.
• If a clip is two seconds short, cover with the next shot or slow it a little rather than regenerating.
• If a clip is the wrong length by a lot, regenerate once with the length set to match. Do not keep spinning.

Step 5: add captions, checked against your script

Automatic captions are a draft. They fail on names, prices, web addresses, numbers and any word that is not common English. If your voice came from your own script, use the script as the caption source instead of transcribing the audio again, then match the timing. If you have no script, run the auto captions and read every line against the audio.

Two kinds of captions exist, and you often want both:

• Burned-in (open) captions are part of the picture. Use them for Reels, TikTok and ads, where many people watch with sound off. ChatCut's guide says Instagram and TikTok do not document caption file uploads, which is why burned-in is the safe choice there.
• A caption file (usually SRT, a plain text file with timings) is uploaded alongside the video. Use it on YouTube and your own site. Viewers can switch it off, and it is better for accessibility.

For how to make and edit those files, see subtitles and SRT files.

On style, the Described and Captioned Media Program (DCMP) Captioning Key prefers mixed case over all capitals for readability, and says sound effects should be captioned when they are needed to understand the video. The Web Content Accessibility Guidelines (WCAG 2.2, the standard behind most accessibility rules) ask for captions on prerecorded video with sound, covering speech, who is speaking and meaningful non-speech sounds, including music. If your video is for a public body or you serve people who rely on captions, treat that as a baseline, not a bonus.

Step 6: safe zones and the sound-off check

A safe zone is the part of the frame the platform's buttons, names and captions do not cover. Meta's ad specifications for Instagram Reels ads have been reported to list safe zones of about 14 percent at the top, 35 percent at the bottom and 6 percent at each side, on a 9:16 picture. The figures are not confirmed on Meta's page as of 4 November 2026, so treat them as a cautious guide and check Meta's current ad specs page. That means the default caption position at the bottom of the frame is the worst place in a vertical ad.

Where	Rule I use
Reels ads, 9:16	Captions in the middle third. Nothing important in the bottom 35 percent or top 14 percent.
Reels and TikTok organic posts	Same middle-third habit. The interface covers the lower part there too.
Feed video, 4:5	Lower third is usually fine, with a margin from the edge.
YouTube, 16:9	Use the caption file, so viewers control size and position.

Then run the check from the exported file, not from the editor preview. Watch it once with sound on, once with sound off on a phone. With sound off, can you follow the point from the captions and pictures alone? With sound on, is every word of the voice still clear over the music? A caption guide from ChatCut (a video editing tool) also advises keeping captions in the middle third of vertical video, away from faces and the platform's buttons.

Free tools and what they limit

You do not need a paid AI service to do steps 3 to 6. A free editor is often the right tool.

• CapCut (a video editor from ByteDance) has a free tier and a paid Pro plan, and its pricing page (read 4 November 2026) offers a 7-day Pro trial that may vary by region. The page does not state commercial-use or music-licence terms, so do not assume a track in its library may go in a client ad.
• OpusClip (a tool that makes short clips and captions) lists a free plan with a monthly credit allowance (60 on one part of its page, 90 on a banner) and a watermark on free clips. Its page gives different language counts in different places, and the 95 percent accuracy figure it advertises is its own claim. A watermarked clip is not usable for client work.
• A plain editor with an SRT import is slower but has no watermark and no surprises.

If one clip needs captions and music, a free editor beats a credit-based AI service on cost and speed. Use the AI service for the pictures, not the finishing.

Where Higgsfield fits for voice and lip sync

Higgsfield is useful here for three things: making the clips, making a voice in the same place, and lip sync when a character must speak your line. Its audio tab covers text to speech, voice change, translation and a custom voice from a short sample, and Lipsync Studio then applies the audio to a face. Doing the picture and the voice in one account saves moving files around.

It is a weak fit for the finishing work. It is not a full editor, so you still need a timeline for music, levels and captions. Credit costs for several of its video features are not published and show on the button, so measure the cost of an approved clip rather than guessing. The plan names and prices change, so check how Higgsfield pricing works.

Disclosure and consent for AI voices

This is information, not legal advice. Check local law and each platform's current rules.

• TikTok has published rules asking creators to label realistic AI-generated content. The current wording is not confirmed on its help page as of 4 November 2026, so read TikTok's AI-generated content page before you post.
• Meta applies its own "AI info" label from signals in the file, and has a separate disclosure rule for altered media in social issue, election and political ads (check Meta's current policy page). See the platform label rules.
• EU AI Act, Article 50: the Act's transparency rules for deepfake audio, image or video (disclose that it is artificially generated or manipulated) were scheduled to apply from 2 August 2026. Check the official text for the current date and scope. A synthetic voice reading your own script is not a deepfake of a real person, but a cloned voice of a real person can be. Check the official text for your case.
• Do not strip provenance data. See why you should leave C2PA marks in place.

Higgsfield's terms also say you must hold the rights and consents for any voice or likeness you put in, and must not present output as human-made when it is not.

Who should skip Higgsfield, and the AI voice

• Skip Higgsfield if you already have the footage and only need captions and music. A free editor does it.
• Skip Higgsfield if your video is mostly a person talking to camera. Film it, caption it and move on.
• Skip the AI voice if the speaker is the trust signal: a founder message, a patient or client testimonial, a regulated claim. Use a real recording.
• Skip the AI voice if you cannot show consent for a cloned voice.
• Skip generated speech inside clips if exact wording matters, such as prices, legal lines or phone numbers. Put those in captions and your own audio.

How I'd approach this

I would do one short video the slow way first: script, voice, music, clips, then captions from the script, noting the music source and consent for each. Time how long each step takes and how many clips you threw away. That tells you whether the order saves you money. Then I would write the order and the checklist above into one page your team follows every time, so a new person does not start by cutting clips.

If you make these often, Blur Studio can build the repeatable workflow: a shared folder for approved voices, licensed music and caption files, plus an approval step before anything is posted. The simplest version is a shared sheet. The tools matter less than keeping the order.

If you do want one place to make the clips and the voice, try Higgsfield on a small project before committing to a plan, and keep the finishing in a free editor.