Silent generate, then TTS, then captions

When the line must be exact, generate silent, speak it with TTS, then burn captions. On this catalog that stack is Kling 3 Turbo (12 credits, silent) → Gemini 3.1 Flash TTS (4 credits, inline audio tags) → caption burn-in. Native audio is convenient and baked. It is the wrong first spend when legal, the founder, or the offer needs those words and no cousins.

This is the sequence. It is not the radio essay about when to pick Turbo versus a sounded row.

Generate silent on purpose

Kling 3 Turbo text-to-video: audio flag off, 1080p, 3–15 seconds, 12 credits, discounted. It will not speak. Plan for that. A quoted line in the Turbo prompt is wasted tokens. You will get motion. You will not get the sentence.

Use Turbo (or another silent card) when:

• The copy is a contract: price, claim, ingredient, disclaimer.
• You will localise the voice later (one picture, many languages).
• You are testing blocking and will throw most takes away.
• You do not want a baked music bed fighting the VO.

Do not use Turbo when you wanted the model’s own performance — room tone, a laugh, lips that invent a line you will never type. That is native audio on Veo 3.1 (40cr), Seedance 2.5 (47cr), Kling O3 Pro (70cr), and neighbours. Baked speech is a take, not a script.

Procedure for the silent plate:

1. Lock ratio (9:16 if it ships vertical) and duration to one beat.
2. Prompt picture only: who, where, the verb, the camera. No dialogue in quotes.
3. If a SKU is on screen, attach a locked still and use I2V, even if that row is not Turbo. Silent T2V will invent a cousin bottle.
4. Generate. Watch muted. You are scoring motion.
5. If the mouth is hanging open as if it wanted to talk, that is fine. TTS does not need a closed mouth if you will not lipsync. If you will lipsync, prefer a face that can take a mouth pass (see video-to lipsync at 4–7 credits) rather than a profile with no lips.

Keep the plate silent until the picture is the keeper. Adding TTS to a miss makes a miss you have now heard twice.

Wan 2.7 I2V is also audio-null on this catalog (15cr, 1080p). It is a silent plate with a still attached. Same sequence: picture, then TTS, then captions.

TTS the exact line, with tags if you need them

Gemini 3.1 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 the voice you already picked. This page uses Flash TTS as the cheap, tagged default.

Procedure:

1. Write the line as it must ship. Not a paraphrase of the brief. The words.
2. Add tags only where a human would actually do that thing. [whispering] ships in twelve days is a choice. [laughing] on a medical claim is a miss.
3. Pick a voice and keep it for the campaign. Six voices in six hooks is not a test.
4. Generate the audio (4 credits). Listen on a phone speaker, not only studio headphones.
5. Check duration against the plate. A 6-second VO on a 4-second clip will overflow. Trim copy or generate a longer plate. Do not speed the voice to 1.4× until it sounds like a disclaimer.
6. Lay the TTS under the silent plate. If you need the mouth to match, run a video-to lipsync row (Kling Lipsync 7cr or VEED Lipsync 4cr) with plate + TTS. If the face is not talking on camera (B-roll, pack orbit, hands), skip lipsync. Lay the voice under. That is enough.

Native audio is wasted on this job because replacing a wrong baked line is an edit, often a regenerate. Four credits of TTS on a 12-credit plate is a 16-credit stack you can recast without touching the picture. A 40-credit Veo take with the wrong sentence is 40 credits plus a second 40, or a messy under-dub.

When native audio is not wasted: you want the model’s performance and you will accept its wording. A cinematic 8-second Veo with room tone is that job. A “20% off until Sunday” UGC line is this page.

Dub video is for an existing language you are replacing. TTS is for a line that never existed on the plate. Do not dub silence.

Caption the TTS, not the plate’s imaginary speech

Burn captions last, from the audio you actually laid in.

Procedure:

1. Picture lock + audio lock. No more Turbo regenerates.
2. Transcribe the TTS (165 language codes on add_veed_captions). You wrote the line, so this pass should be clean — still spot-check names and numbers.
3. Pick BASIC (21 looks, 1×) or DYNAMIC (9 looks, 2×). One family for the campaign.
4. Place in the safe area.
5. Burn via add captions to video.
6. 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 plate before TTS. There is nothing to transcribe. Do not prompt “yellow captions” into Turbo. That is texture.

If you localise: new TTS in the new language, then new captions from that audio. The silent plate can stay. That reuse is the reason you did not bake English into the generate.

Stack recap in credits (catalog, not USD):

Step	Row	Credits
Silent plate	Kling 3 Turbo T2V	12
Voice	Gemini 3.1 Flash TTS	4
Mouth (optional)	VEED Lipsync / Kling Lipsync	4 / 7
Captions	BASIC 1× or DYNAMIC 2×	render tier

Compare that to one native-audio hero you might still run after the line is proven. The sequence is how you get to that hero without paying for six wrong sentences.

If you need a studio seat to run Turbo, Flash TTS, and the burn-in, sign up. Bring the exact line. Do not bring “make it sound natural” without the words.

FAQ

Why not native audio and then captions?

You can, when the model’s wording is acceptable. When the line must be exact, baked speech is a liability. Silent + TTS lets you change the sentence without regenerating the picture.

Do I always lipsync the TTS onto Turbo?

No. Lipsync if there is a mouth the viewer can see talking. Pack B-roll and hands do not need it. Video-to lipsync is 4–7 credits extra for a reason.

Can I caption before TTS?

No. Captions follow speech. Silent plates have none. TTS first, then transcribe and burn.