Word-highlight needs clean speech

Word-highlight needs clean speech. Karaoke captions light each token as it is spoken. That clock is a word timestamp, not a font. Music under the voice, two people at once, and a loud whoosh will throw the accent onto the wrong syllable. Isolate the vocals, or drop to a static BASIC block. Do not buy a 2× DYNAMIC render of a dirty mix.

The clock is per word, not per vibe

A phrase-timed cue puts the whole line up and takes the whole line down. Karaoke keeps the line, and a clock walks through it: the current token takes the accent, the rest wait. Forced alignment is how you get the clock from a known transcript. Speech-to-text is how you get the words. Burn-in is how the look survives a muted feed.

On Caption Studio, Snap (effect=pop) and Reveal (effect=typewriter) are the timed pair in the 45 CSS presets (9 groups × 5). On the VEED burn, the 9 DYNAMIC looks (glass, whisper, glide, fusion, terminal, handwritten, backdrop, and the rest) are word-level, billed at 2×. 21 BASIC looks are line-level at 1×. Same transcript. Different clock.

If the aligner cannot hear a word, it cannot stamp it. A late stamp is a visible miss in a way a slightly late block is not. The viewer is watching the highlight, not just the sentence. That is the point of karaoke, and why a dirty mix looks broken instead of merely imperfect.

This is not the page that says karaoke is word-highlight. That page exists. This page is the audio condition: clean speech, or do not use the clock.

A silent product plate has no word to time. An authored overlay nobody said is not karaoke. Apply the timed pass on speech. Keep the active word off the mouth and off the chrome.

Music-under and overlap are the usual wreck

ASR and alignment assume a voice in front of a bed, not a voice inside a bed.

Music under. A bed that sits at the same loudness as the talker is a competing fundamental. The aligner will stamp drums as syllables or drag the accent toward a downbeat. Suno and other music rows are fine as beds. They are not fine as equal partners with the VO if you need word timestamps. Dip the bed 12–18 dB under the voice for the karaoke pass, or export a vocal stem and align to that, then burn onto the mixed picture.

Overlap. Two people, a street interview, a webinar with no isolate. The clock cannot serve two mouths. Speaker IDs are a BASIC problem. Karaoke on overlap is two timestamps fighting. Split the clip, or drop to a block, or author the line.

SFX hits. A whoosh or a door on the same frame as a plosive will steal the stamp. If the door is the joke, SDH as a static label is honest. Animating the whoosh as if it were a word is not.

Reverb and generate mix. Native-audio models bake room into the take. A Veo line in a cathedral is harder to stamp than a dry clone. If karaoke is the creative, prefer a dry VO mixed after, or accept BASIC on the wet bake.

Compression and phone speakers. Alignment runs on the file you upload, not on the speaker the scroller owns. The wreck happens at mix, not at playback. Clean the stem before you caption.

Stem separation is a tool, not a miracle. If the bed was generated into the same native-audio take as the words, there may be no stem. Then you do not have a karaoke job. You have a BASIC job, or a regenerate with VO separate.

Caption timing complaints on Snap are often this, blamed on the preset. Recolouring the accent will not fix a stamp that landed on the snare.

Isolate, then timestamp, then burn

Order:

1. Lock the picture.
2. Lock a vocal-forward mix, or a vocal stem. If you cannot, skip karaoke.
3. Transcribe. 165 language codes. Code matches the talker, not the song’s language.
4. Read names and numbers. A highlighted wrong SKU is still wrong.
5. Generate word timestamps on the clean speech.
6. Pick Snap or Reveal, or a VEED DYNAMIC look. Preview on this plate. Contrast still has to hold. Glass on a busy pack will vanish even when the stamps are perfect.
7. Watch on mute at phone size. If the accent trails the mouth, that is alignment, not colour. If you cannot finish the line, that is CPS, not karaoke. Split the cue. Do not add motion to a maxed line.
8. Burn. 2× on DYNAMIC. Do not buy the multiplier on a dirty mix.

The AI caption generator and add captions to a video will apply a timed look to whatever audio they are given. They will not solo your vocals for you. Bring the clean speech.

If the organic hook needs karaoke and the paid twin needs a freeze-frame, same transcript, two burns: DYNAMIC on the clean stem for organic; BASIC on the same words for paid. Do not ship the dirty-mix Snap into Ads Manager.

SubRip cannot represent per-word highlight. If a client asks for karaoke in an .srt, they want a burned MP4 plus a plain sidecar for archive, or they have the wrong extension in the brief. Convert the archive with the VTT to SRT converter. Burn the look.

When to quit karaoke

Quit when:

• the bed cannot be dipped and there is no stem
• two speakers overlap on the hook
• CPS is already maxed on a static block
• the cut is a jump-cut montage (the clock cannot survive 12-frame shots)
• the take is a silent demo with authored type

A two-line BASIC block is not a failure of taste. It is the honest clock when word timestamps would be fiction. Fiction karaoke looks like a broken lyric video. Reviewers pause on it. Scrollers feel it as lag.

A month of organic hooks on Glass is a month of 2× renders. Spend that multiplier on a vocal-forward mix you aligned, not on a native-audio bake with a bed you cannot solo. The cheap fail is BASIC on a dirty take. The expensive fail is DYNAMIC on the same take.

Word-highlight needs clean speech. Isolate the voice. Then stamp. Then burn.

Align the next karaoke pass in Versely.

FAQ

Can I leave the music loud and still use Snap?

You can. The accent will drift toward the bed. Dip the music for the align, or drop to BASIC. 2× DYNAMIC will not hear through a loud mix.

Is a native-audio Veo take clean enough for karaoke?

Sometimes, if the line is dry and solo. Wet rooms and baked music are not. Preview the stamps. If the pop misses the mouth, BASIC the bake or replace with a dry VO.

Does Reveal survive a dirty mix better than Snap?

No. Both need word timestamps. Reveal’s leftover words can hide a missed pop; they cannot hide a clock that is on the snare. Clean speech, or a static block.