Replace the take's audio after TTS

Generate silent. Speak the line with Gemini 3.1 Flash TTS (4 credits). Then replace the video’s audio. Replace mode mutes the original and lays your URL in its place. Mix is the other door, and it is wrong when the plate was silent on purpose or when the leftover native take would fight the read.

This is the attach step after TTS. It is not the essay on when to pick Turbo versus a native-audio row.

Generate the plate silent on purpose

If you needed those exact words, do not buy a baked performance and then throw it away.

Silent plate rules:

Pick a silent card. Kling 3 Turbo is 12 credits, audio false, 1080p, 3–15s. Wan 2.7 image-to-video is audio-null on this catalog (15 credits, 1080p). Quoted dialogue in a silent prompt is wasted tokens.
Lock ratio and duration to the line you will speak. A 4-second plate under a 6-second VO will overflow. Write the line first, time it out loud, then set duration.
Prompt picture only: who, where, the verb, the camera. No quotes.
If a SKU is on screen, attach the locked still and use image-to-video. Silent text-to-video will invent a cousin bottle.
Generate. Watch muted. You are scoring motion. Mouth hanging open is fine if you will not lipsync. If you will lipsync, prefer a face that can take a mouth pass.
Sign the picture. Do not attach TTS to a miss. A miss you have now heard twice is still a miss.

Native audio is the other stack: Veo 3.1 at 40 credits with a baked line. Use that when you want the model’s performance and will accept its wording. This page is the contract line: price, claim, ingredient, disclaimer, founder script.

Keep the silent parent. Replace is a child that throws the original soundtrack away on purpose. On a silent plate that soundtrack is empty or nearly empty, which is why replace is the honest mode.

TTS the exact line, then check the clock

Flash TTS is the 4-credit speech row: 30 voices, natural-language style control, inline audio tags ([sigh], [laughing], [whispering]), multilingual, multi-speaker dialogue. Cartesia Sonic 3.5 is 5 credits if that is already the voice. This procedure uses Flash TTS.

Procedure:

Paste the shipping sentence. Not a paraphrase of the brief.
Add tags only where a human would do that thing. [whispering] on a medical claim is a miss. [laughing] on a price is a miss unless the brief is a joke.
Pick one voice and keep it for the campaign. Six voices in six hooks is not a test.
Generate the wav (4 credits). Listen on a phone speaker.
Measure duration against the plate. If VO is longer, cut copy or generate a longer plate. Do not time-stretch the voice to 1.4× until it sounds like a disclaimer.
If the mouth must match, run video-to-lipsync on plate + TTS before you think about replace — the lipsync row already writes a new file with new audio. Replace is for when the picture stays and only the bed changes.
Keep the TTS URL. attach_audio_to_video needs an HTTPS file, not a local path.

Checkable audio QC:

Names, numbers, and the SKU pronunciation.
Dead air at the head of the wav. Trim the wav if the model left a handle; do not leave 400ms of silence that makes the picture feel late.
Loudness versus a phone speaker, not only studio headphones.

The AI video generator made the plate. TTS is a speech row. Neither of them attaches the files. That is the next section.

Replace, do not mix

attach_audio_to_video in mode replace strips the video’s original audio entirely and substitutes the file you supply. Mode mix keeps some of the original. On a silent plate there is nothing worth keeping. On a native-audio plate you decided to throw away, mix will leave a ghost of the old line under the new one.

Procedure:

Confirm picture lock. Replace on a candidate is a mute you will repeat when you regen.
Point replace-video-audio at the silent (or condemned) video URL and the TTS URL.
Set mode to replace. Do not “mix at 5% original” as a hedge. That hedge is the old take in the noise floor.
If the wav is shorter than the video, decide: trim_to the wav length, or hold picture on the tail in silence. Silence on a CTA hold is allowed. Silence under a still-moving pour looks like a mistake.
If the wav is longer, you already failed the clock check. Cut copy or extend the plate. Do not let replace spill.
Render. The video stream is copied; only the soundtrack changes. That is the point. You are not re-encoding the pack.
Listen to the child from the first frame. Confirm the original is gone. Confirm the first word lands where you wanted.

When replace is the wrong tool:

You wanted room tone under the VO. That is mix with a low original_volume, and only if the original is actually room, not a competing line.
You wanted a new language of an existing performance. That is dub video, not TTS-plus-replace.
You wanted the mouth to match. Lipsync first. The lipsync output already carries the audio.

Do not replace, then regen the picture, then replace again. Two attaches. The first taught you nothing. Lock picture, then attach.

Caption the new soundtrack, not the empty plate

Burn last, from the audio you actually laid in.

Procedure:

Picture lock + audio lock on the child.
Transcribe the TTS. You wrote the line, so the pass should be clean — still spot-check names and numbers.
Pick BASIC (21 looks, 1×) or DYNAMIC (9 looks, 2×). One family for the campaign.
Place in the safe area.
Burn via add captions to video. 165 language codes. Language is what is spoken now, not what you wished.
Watch muted. The caption must match the TTS. If you “tightened” the caption into a different claim, you now have two scripts.

Do not transcribe the silent parent. There is nothing to transcribe. Do not leave captions you burned on a native-audio take you then replaced; those cues are a different performance.

If you localise: new TTS in the new language, replace again (or replace onto a copy of the silent parent), then new captions. The silent plate can stay. That reuse is why you did not bake English into the generate.

Stack recap in credits, catalog not USD:

Step	Row / tool	Credits
Silent plate	Kling 3 Turbo T2V	12
Speech	Gemini 3.1 Flash TTS	4
Attach	replace-video-audio	editor fee (flat, length-agnostic on this tool)
Captions	add_veed_captions	1× BASIC or 2× DYNAMIC

Lipsync, if you added it, is extra (VEED 4cr or Kling 7cr) and usually replaces the need for a separate attach.

Lay the TTS onto the silent keeper in Versely. Bring the wav URL and the plate. Do not bring a mix percentage.

FAQ

Can I mix TTS under native audio to keep the room?

Only if the native take has no competing line. A leftover baked sentence under TTS is two scripts. Silent generate, then replace, is the clean version of this stack.

Does replace re-encode the picture?

No. The video stream is copied. Only the soundtrack changes.

Should I caption the silent plate before TTS?

No. Caption the replaced child, from the audio you laid in.