Word-by-Word Captions for Clips: The Style That Holds Attention

Word-by-word captions show two to four words at a time and mark the word being spoken, by color, a karaoke fill, or a fade-in. To make them, transcribe the clip with word-level timestamps, group the words into short cues, burn them into the video, and place them where nothing important sits in the vertical frame.

Many people scroll with the sound off, and a caption that moves with the voice gives the eye something to follow. For clippers it is also a choice fully in your hands: the campaign picks the footage, you pick how it reads. This guide covers the settings that matter and how Treza's AI video pipeline burns them. New to paid clipping? Start with how paid clipping campaigns work.

What makes a caption "word by word"

Plain subtitles show a full sentence for as long as it is spoken. Word-by-word captions change three things:

• Short cues. Two to four words on screen at once instead of a sentence. The text changes as fast as the speaker talks, so it reads like speech, not like a transcript.
• A marked spoken word. The word being said stands out: a different color, a color that fills it left to right, or a fade-in that builds the line as it is spoken.
• Heavy type. A thick condensed or rounded display face with an outline, big enough to read on a phone at arm's length, often in capitals.

All three depend on something a basic subtitle file does not carry: when each word starts and ends. So this look starts from a transcript with word-level timestamps, not an SRT. For subtitle files versus burned-in text in general, see burned-in captions vs SRT files; for the transcription side, how to add captions automatically. This post is only about the word-timed, burned-in style.

The settings that matter

The names below are the Captions node's settings in Treza, as of November 2026.

Words per caption

The biggest choice. The node takes 1 to 12 words per caption, or 0 for full sentences, and defaults to 4.

• 1 word is the loudest look. It suits hype edits and gets tiring over a 60-second story.
• 2 to 3 words is the common talking-head setting. Each cue lasts about a second for most speakers, so the eye never has to scan.
• 4 to 6 words reads calmer and suits slower speakers, explainers and interviews.

The Campaign Clip template ships at 3.

Spoken word highlight

Beyond no highlight at all:

• Color the spoken word. The current word switches to the highlight color. This is the classic podcast-clip look, and the template default (yellow).
• Karaoke fill. The spoken word fills with color from left to right as it is said, like a lyric video.
• Reveal. Every word keeps the same color and fades in as it is spoken, so the line builds up word by word. Quieter, closer to a documentary.

Highlight colors are yellow, orange, lime, cyan, pink and white. If you pick the same color as the caption text, the node shows the spoken word in yellow instead, so the highlight never disappears.

Style, font and case

Four styles: bold (thick outline), clean (soft shadow), boxed (text on a dark box), and plain (no outline, for solid backgrounds). Five fonts: a default sans, Anton (tall condensed), Archivo Black (wide heavy), Montserrat ExtraBold (rounded geometric) and Inter SemiBold (plain editorial). An ALL CAPS switch sets the text in capitals.

Bold with a display face is safest on busy footage, since the outline keeps white text readable over a bright background.

Motion and profanity

Each caption can enter with a pop, a fade, or a brief red and cyan glitch, and punch-ins zoom the picture on caption changes, the jump-cut rhythm of talking heads. Swear words are masked after the first letter by default (f***), which also catches speech-to-text mishears that would put a word on screen nobody said.

A color per speaker for podcast two-shots

Two voices in one caption track blur together. Color by speaker fixes that: the caption changes color when the voice changes.

Run the Transcribe node with Label speakers on, which tags every word with a speaker ranked by talk time, then pick what changes in the Captions node's Color by speaker setting:

• Caption text recolors the whole caption for each voice.
• Spoken word only keeps the text color and changes the highlight per voice.
• Text and spoken word does both, and is the one to pick with a highlight on, because each voice keeps its own spoken-word color and a one-word interjection still reads as a change.

The speaker who talks most keeps the colors you set, and every other voice takes the next color along. Speakers to distinguish defaults to 3 and goes up to 6; voices past it share the last color, so a diarizer that hears three voices in a two-host show cannot add a color nobody chose. The Interview preset sets this up in one click: clean text, Inter, a color highlight, and Text and spoken word.

Where the captions go on a vertical frame

Placement depends on how the clip was reframed. The node has five positions: top, upper middle, center, lower middle (the default) and bottom.

• Subject-centered 9:16 crop. The speaker fills the frame, so captions at lower middle sit below the face and above the platform's own buttons and caption text. Center works for the big one-word and podcast-punch looks, as long as it does not cover the mouth.
• Blurred canvas. When Extract Clip puts a square window, or the whole 16:9 frame, on a blurred and darkened copy of the picture, the blurred band under the picture is a clean place for text. Set Position to Bottom and the captions land in the band, off the faces. The horizontal to vertical guide explains when to pick each reframe.
• Busy lower third. If the source already has on-screen text low in the frame, move the captions to upper middle rather than stacking text on text.

Preview on a phone either way: the account name, post caption and buttons cover the bottom of a Shorts or TikTok frame.

Presets, if you would rather pick a look

The presets row applies a canned set of settings, and every dropdown then shows what changed so you can fine-tune. A few that suit clipping:

Preset	What it sets	Good for
Podcast punch	Anton capitals, centered, 3 words, yellow spoken word, pop, snap punch-ins	Talking-head Shorts
Karaoke fill	Montserrat capitals, centered, cyan karaoke fill, pop	Music, hype, lyric-style
Interview	Clean Inter, lower middle, 5 words, color per voice	Podcast two-shots
One word	Anton capitals, one word per beat, centered	Short punchy lines
Documentary reveal	Clean Inter, 6 words, reveal, fade	Slower stories, explainers

How it runs in a clipping pipeline

In Treza's Campaign Clip template, the campaign brief and the source video go in. A language model picks a moment that fits the brief, Extract Clip cuts and reframes it to 9:16, and the clip is transcribed again on its own timeline with Scribe v3, so word timings land on the cut instead of drifting from the source. The Captions node burns them in: bold, lower middle, three words, the spoken word in yellow, Montserrat. Change that once on the node and every later run uses it.

A judge model then watches the captioned clip with the brief and scores it on brief fit, the hook, length, caption accuracy, and banned or required on-screen content. Under 70, or on anything the brief forbids, the run stops before it posts.

The template posts to YouTube Shorts, and to TikTok with a TikTok Upload node added. For Instagram Reels and X you get the finished MP4 and upload it by hand. For captioned vertical clips from your own long video without a brief, see the AI clip generator. When a campaign wants original footage rather than clips, the AI video generator is the separate tool.

Treza is independent of every marketplace named here. Clip only the footage a campaign or the rights holder allows you to use.

What campaigns say about captions

As of November 2026, the marketplace pages we checked set no caption rules of their own. Vyro's content requirements ask clippers to disclose their connection to the client and to post only their own content. Content Rewards' help pages say a submitted post must be public and under 30 minutes old. So caption rules, when there are any, live in each campaign's brief. Read it before you pick a look; why clips get rejected covers the rules that do get clips turned down.

One platform note: TikTok's Community Guidelines (as of November 2026) keep "unoriginal or reused material without anything new" off the For You feed. We cannot say how TikTok weighs a caption track, so do not count on captions alone to make a clip new.

Frequently Asked Questions

How do I make word-by-word captions?

Transcribe the clip with a model that returns word-level timestamps (Whisper or ElevenLabs Scribe v3), set words per caption to 2 or 3, turn on a spoken-word highlight, and burn the captions into the video. In Treza that is a Transcribe node wired into a Captions node, and the Campaign Clip template has both wired already.

How do I get TikTok-style captions on a clip?

Bold style with a heavy display font (Anton, Archivo Black or Montserrat), capitals, two or three words per caption, and the spoken word in yellow, placed at lower middle or center.

What are karaoke captions?

Captions where the spoken word fills with color from left to right as it is said, like a lyric video. The Karaoke fill preset sets it up with centered capitals.

How many words per caption should a clip use?

Two to three for most talking-head clips, one for short hype lines, four to six for slower speakers and interviews. If a fast talker makes the text flicker, add a word.

Can captions change color for each speaker?

Yes. Run the Transcribe node with Label speakers on, then set Color by speaker on the Captions node. Pick Text and spoken word if you also use a highlight, so each voice keeps its own spoken-word color.

Do clipping campaigns require captions?

The marketplace pages we read (Vyro and Content Rewards, as of November 2026) do not require them; individual campaign briefs might. Check the brief, and caption anyway: a clip watched on mute still reads.

Do burned-in captions work on Instagram Reels and X?

Yes, they are part of the picture. Treza posts to YouTube Shorts and TikTok; for Reels and X, upload the finished MP4 by hand.