Isolate vocals before karaoke captions

Isolate the vocal, then burn karaoke captions. A dirty mix — music bed, whoosh, a second talker, room slap — makes word-highlight land on the wrong syllable. Phrase-level BASIC looks forgive a little drift because the whole line is up at once. Karaoke does not. Isolate vocals from a track (isolate_audio) is the general-purpose strip. Caption the clean speech, not the bed.

This is the audio gate before a word-timed burn. It is not a Snap-versus-Glass style tour.

Hear why the mix will lie to the word clock

Karaoke (word-timed captions, Snap/Reveal families, VEED DYNAMIC looks such as glass, whisper, glide, fusion) consumes a start time per token. If the recognizer hears the snare as a consonant, or smears a lyric into the next word, the accent walks onto the wrong syllable. The viewer who cannot hear the file thinks the edit is drunk. The viewer who can hear it sees the highlight miss the mouth.

Dirty sources that fail this clock:

UGC with a loud music bed under the line.
A native-audio generate that also baked a score.
A podcast clip with two mouths and a theme.
A karaoke-the-song job where you wanted lyrics on the vocal, not on the instrumental.

Procedure before you touch captions:

Listen on headphones, then on a phone speaker. If you cannot transcribe the line by ear in one pass, neither will the engine.
Note competing events on a pad: 0:00–0:04 bed, 0:02 whoosh, 0:06 second voice.
Decide whether the job is speech captions (UGC line) or lyric karaoke (the song is the point). Both want a vocal isolate. They do not want the same later mix.
If the file is already a dry VO, skip isolate. Do not strip a clean mic and hope.
If the file is MOS (no speech), stop. Karaoke has no word to time. Overlay is the other object.

BASIC at 1× is the fallback when you cannot isolate and still need words on screen. Shipping DYNAMIC at 2× on a muddy bed is how you pay extra to advertise the drift.

The AI caption generator will run on whatever audio is in the file. It does not know you meant “ignore the 808.” Isolation is your gate, not a preset.

Isolate the vocal, then decide what the picture hears

isolate_audio takes any mixed audio_url — not only in-app Suno tracks — and returns a vocal with background music and instrumentation removed. Quality follows the mix. A clean dry-ish recording separates. A smashed TikTok master separates less. For a quick split on your own device, the free vocal remover separates vocals from the instrumental in your browser, with no upload.

Procedure:

Pull audio from the video if you only have a muxed file, or point at the wav you already have.
Run isolate vocals from a track.
Listen to the isolate at 1×. Check names, numbers, the SKU, and the ends of words. A chopped sibilant will still throw a timestamp.
If the isolate is a smear, you do not have a karaoke source. Either re-record / TTS a dry line, or burn BASIC on the original and live with phrase-level timing.
Keep the isolate URL. You will need it twice if you also recaption from it.

Two later doors. Pick one on purpose.

Door A — caption from the isolate, ship the original mix. You wanted the bed in the ad. You did not want the bed in the transcript. Transcribe / word-time against the isolate, then burn those cues onto the original picture. This is the usual UGC case: music stays, highlight follows the voice.

Door B — replace the video’s audio with the isolate, then caption. You wanted a dry line: a recut, a lyric video, a VO you will dub later. Replace video audio the isolate onto the picture, then caption that child.

Do not confuse this tool with separate_music_vocals, which needs a prior generate_music taskId/audioId. That is the Suno-stem door. This page is the general mix.

If the competing sound is a second talker, isolation will not give you a clean speaker split. That is a two-person ID problem, not a stem problem. Label speakers after a clean as you can get transcript; do not expect karaoke to un-overlap two mouths.

Burn karaoke only after the word clock is honest

Word-timed looks last.

Procedure:

Lock picture. A later trim is a new clock.
Point the caption job at the audio you trust: the isolate (door A, if your caption path can take a sidecar transcript) or the replaced child (door B). On Versely, add captions to video transcribes the video’s own soundtrack. If you need the engine to hear the dry vocal, door B is the honest path: replace, then caption. If you must keep the bed in the file, isolate still helps you QC the words and correct the transcript before you burn; a muddy auto pass will still guess.
Correct the transcript by hand: names, numbers, SKU, the claim. Wrong words with perfect timing are still wrong.
Pick a word-timed look only now: DYNAMIC (9 looks, 2×) or a Snap/Reveal family in caption styles. BASIC (21 looks, 1×) if the clock is still shaky.
Preview on the first seconds. preview_caption_style exists so you do not buy a full DYNAMIC render to learn the highlight is late.
Burn. Watch muted at phone size. The accent must travel with the mouth. If it sits on the previous word, you are back to audio, not to a punchier preset.
Recheck safe area. Karaoke that crawls onto the mouth or the chrome is two problems.

Checkable mute test:

First stressed word of the hook: does the highlight land with the viseme.
A short function word (“the”, “a”) next to a long noun: does “the” steal the accent from the noun. Interpolation inside a segment does that. Isolation will not fully fix interpolated word times, but a dirty bed makes interpolation worse.
Last word of the line: still up long enough to read. Karaoke can feel faster than it is.

If the accent is still drunk on a dry isolate, the timestamps were interpolated inside a segment. Restyling will not fix that. Split the cue, hold the line as BASIC, or accept phrase-level type. Do not stack a second DYNAMIC preset as a repair.

Put the bed back only if the brief still wants it

Isolation is a working file. The ship mix is a decision.

Checklist:

If the ad needs energy from the bed, keep the original soundtrack in the picture file and live with door A’s limits, or mix a lower bed under a dry VO you control (TTS) instead of captioning through a smash.
If you replaced with the isolate, you may mix a new bed after captions only if you are not going to recaption. Captions already burned will not care. A later recut will.
Do not isolate, caption, then slap the original smash back on without checking that the words still match. You can drift a frame in the attach.
Lyrics jobs: the isolate is the source of truth for words. The instrumental can return under, quieter. Highlight still follows the vocal you captioned.
Archive the isolate next to the video. A recaption in a second language needs a new dry read, not this English stem.

Karaoke on a dirty mix is how “ships in twelve days” highlights “twelve” while the mouth is still on “ships.” Strip first.

Isolate, then burn the word clock in Versely. Bring the mixed URL. Do not bring a DYNAMIC preset as the first step.

FAQ

Can I just pick BASIC and skip isolation?

Yes, if you only need a phrase-level block. BASIC forgives more drift. Karaoke and DYNAMIC do not. If the look is word-timed, isolate first.

Does isolate_audio work on any file, or only Suno?

Any mixed audio URL. The Suno-only stem split is a different tool and needs a prior generate_music id.

Why not fix a late highlight by switching from Glass to Whisper?

Both follow the same word clock. A new preset restyles the lie. Clean the audio or drop to a phrase-level look.