Add voiceover after a silent generate

A silent row returns picture. Write the script, generate speech, then add a voiceover to the video. Replace the empty track. Do not prompt quoted dialogue on Kling 3 Turbo (12 credits, audio false) and treat the mute export as a defect. The catalog already said the soundtrack was yours.

This is the attach after a silent generate you already have. It is not the radio essay on when to pick Turbo versus native audio, and it is not the replace-mode footnote after TTS.

Read the audio flag before you prompt a line

The catalog stores audio as true, false, or null. Native audio is true: dialogue, effects, ambience in the same generate. A silent row is false or null. You get an MP4 with no soundtrack, or with a track that is empty. That is the older pipeline: picture, then sound.

Published silent / null examples you may name:

• Kling 3 Turbo text-to-video — 12 credits, audio false, 1080p, 3–15s.
• Wan 2.7 image-to-video — 15 credits, 1080p, 2–15s, audio null on this catalog. Do not borrow native audio from Wan 2.6.
• Gemini Omni Flash Edit — 13 credits, edit-video, audio false.

Native-audio neighbours (wrong first spend when the line is a contract): Veo 3.1 40 credits, 4/6/8s, 4K; Seedance 2.0 69 credits, 4–15s, 1080p. Baked speech is a take. It is not a script you can restyle.

If the file in front of you came from a silent row, do not send it back through a native-audio row “to add sound.” You will get a new picture. Voiceover is how speech gets onto this picture. Lipsync is how a mouth in frame gets matched to a track you already have. Those are later passes.

A brief that says “she says the line” on a silent row will give you a mouth flap and no voice, or a mouth that never moved. Read the row. Then write picture-only prompts: who, where, the verb, the camera. No quotes.

If you already prompted quotes on Turbo, the picture may still be the keeper. Keep it. The unused tokens are gone. The VO still goes on afterwards.

Write the script to the plate's duration

generate_speech turns text into a wav. It is not auto-timed to on-screen actions beyond starting at the beginning of the attach. The plate’s duration is the budget.

Procedure:

1. Watch the silent keeper muted. Confirm it is a keeper. A miss you then narrate is a miss you have now heard twice.
2. Time the file. 8 seconds is 8 seconds. A paragraph is not an 8-second read.
3. Write the line. Speak it out loud against a timer. Cut words until it fits with a breath. If it cannot fit, you owe a longer plate or a second shot — not a faster talker.
4. Emotion and style instructions exist so the read is not flat. Support varies by TTS model. Gemini 3.1 Flash TTS is 4 credits, inline audio tags ([sigh], [whispering]), multilingual. Cartesia Sonic 3.5 is 5 credits. Stock voice if identity is not the point. Clone if it is, and only with consent.
5. estimate_cost on a long script. Speech bills per generation by model and length. Attach is a separate flat fee.

Mode on attach_audio_to_video:

• replace — narration becomes the only track. Correct on a silent row. There is nothing worth keeping.
• mix — keeps original audio and layers the new file. Correct when the plate has usable room tone you want under the read. A silent row often has no such tone. Mix on digital silence is still just the VO, with extra steps.

If you will lipsync, generate the voice first, then run a mouth pass on a face that can take it. Voiceover on a profile with no lips does not need lipsync. Do not pay for visemes the camera never showed.

Multi-speaker is a different tool. Two aliases belong on add multi-speaker dialogue, not two generate_speech calls you stitch by hand.

Attach replace, not mix, on a silent file

attach_audio_to_video copies the video stream losslessly. Only audio is touched. You are not re-encoding the pack shot by adding a wav.

Replace procedure:

1. Confirm the wav duration versus the plate. A 10-second read on a 6-second plate overflows. Trim the script or extend the picture as a separate job. Extend is new generated frames, not a slow-mo of the last still.
2. Attach, mode replace, VO URL on the silent MP4.
3. Listen on a phone speaker, then on mute. Mute is the feed. The read is for unmute and for captions.
4. If the first second has no picture claim, the VO will not save it. Silent-autoplay still needs a burn or a visual hook. Kapwing’s 2026 UGC notes still treat on-screen words as the mute lever.

Do not mix a music bed in the same attach as the first VO pass unless you already know volumes. Get the voice legal, then mix a bed under it with original_volume high and music_volume low. A bed generated with vocals is a second mouth. Prompt an instrumental.

Native-audio files do not belong on this page. If you accidentally ran Veo, mix a bed under the baked line, or replace only if you are throwing the baked line away on purpose. Throwing away a 40-credit performance to lay a 4-credit TTS is a stack you should have started silent.

Caption the voice you attached

ASR captions transcribe what is on the file now. Caption after attach. Captioning the silent plate produces an empty track or garbage from hiss.

Add captions to a video: 165 language codes, 21 BASIC at 1×, 9 DYNAMIC at 2×. Language must match the VO language. The tool transcribes; it does not translate. If you will localise, keep the silent picture, generate a second VO, replace, recaption. That reuse is why the row was silent.

Burned-in captions are the short-form object. Place them in safe area after the ratio is the ship ratio. YouTube Shorts puts a photoreal AI overlay on the picture (YouTube Blog, 27 May 2026). Leave room.

Spot-check names, prices, and ingredients. TTS will pronounce a SKU the way the model likes. The burn will spell what was said, which may not be the legal string. If the claim must be exact on screen, that string is also an overlay you wrote.

Article 50 (2 August 2026) still applies to photoreal AI at first exposure. A silent row plus a human-sounding TTS does not make the clip “not AI.” Disclose where the destination or the jurisdiction requires it. NY GBL § 396-b (in force 9 June 2026) is a synthetic-performer ad disclosure when the person is not an identifiable natural person — a different object from a voiceover over a pack.

Lay the voice on the silent keeper in Versely. Replace the empty track. Then caption what you attached.

FAQ

The Turbo file has no sound. Did the generate fail?

No. Audio false means picture only. Voiceover is the intended path. Native audio was a different row.

Replace or mix?

Replace on a silent plate. Mix when you are keeping usable original sound under a new bed or a new VO. Digital silence is not usable original sound.

Can I use my own cloned voice?

Yes, if you have the right to. Clone first, pass voice_id into generate_speech, then attach. Consent is the gate, not the credit sticker.