Lay B-roll under an AI voice track

Lay TTS under product image-to-video, then caption. A talking head is optional. The voice is the spine. B-roll is the picture — pack, hands, room, process — cut to the read. Do not start with a generated presenter and then hunt for footage to hide them. Do not prompt a video model to “talk about the product” and wait for a usable mix.

This is an assembly. Voice, picture, type. Three objects. Add B-roll to a video is the edit door. Gemini 3.1 Flash TTS is a voice door. A silent I2V row is a picture door. Captions are last.

Voice first, then picture

Write the read before you generate a frame. B-roll that is more interesting than the sentence is a mistake dressed as a good shot. A sentence with nothing to be true about is a radio ad. The script is the brief for both.

Keep the read short enough to cover. A sixty-second VO with four seconds of product motion is how you loop the same orbit until the viewer notices. One claim per cutaway is a usable unit. If the script has six claims, you owe about six pictures, not one ten-second hero.

Generate the voice on a TTS row, not inside the video prompt. Gemini 3.1 Flash TTS is 12 credits per 1,000 characters: 30 voices, natural-language style control, inline audio tags ([sigh], [laughing], [whispering]), multilingual synthesis, multi-speaker dialogue. Cartesia Sonic 3.5 is 4 credits per 1,000 characters if you need that roster instead. Do not ask a silent I2V row to invent a voice; the catalog audio field on those plates is off or unset.

Lock the take. If you will change the line, change it now. Captioning, lipsync, and picture duration all follow this file. A later rewrite is a new transcript, a new caption burn, and a picture track that no longer matches the clock.

Place the voice on a timeline. Mark the claims. Those marks are where B-roll changes. A cut in the audio is invisible under a cutaway; that is one of the jobs B-roll exists to do. Do not cut the picture on a random beat because a generate came back at 6 seconds.

If the brief needs a human-sounding peer and you have a real recording, use that. TTS is the spine when you do not have a booth, or when you need a second language without a second shoot. 92% of consumers trust a peer over a brand message in the Kapwing 2026 / Billo compilations; a peer can be a pair of hands and a voice. It does not have to be a generated face.

Product I2V as B-roll, not a second host

B-roll under a voice is supporting footage. It illustrates the claim, hides edits, and sets pace. It should not speak. It should not pull focus from the line being read. Motion slower than the speech. No native dialogue. No hard cut inside a four-second plate unless you meant it.

That is why silent I2V is the default picture row for this assembly:

Hailuo 2.3 Fast: image-to-video only, 6s or 10s, 12 credits (was 23), discounted, requires an image. Catalog audio is unset. One still in, a short cinematic clip out. Good for pack steam, a slow push, a turn.
Hailuo 2.3 Pro: 40 credits, 1080p, 6s or 10s, audio false, requires an image. Same duration pair, a 1080p listing. Not a talking upgrade.
Kling 3 Turbo Text to Video: 22 credits (was 27), T2V, silent, 1080p, 3–15s, discounted. Use when you have a sentence and no still. Still silent.

Bring an approved still when the SKU must hold. Crop it to the ratio you will ship before you attach it. A 16:9 pack shot fed to I2V is a landscape plate; you will be on the resize page next. The AI video generator is the run door; the still is the contract.

Prompt motion, not a new commercial. “Slow push in, label stays sharp, steam.” Not “a cinematic ad for the drink with a smiling host.” The host is optional and lives on a different row. The commercial copy is the TTS file.

Length: match the claim, then cut. Hailuo only sells 6s and 10s. A three-second beat still buys 6s; trim in the edit. Trimming does not refund. Two six-second plates for two claims is cleaner than one ten-second plate that invents a second label in the last third.

Stock UGC is the other picture source. Add B-roll to a video is built on get_available_broll, search_broll, and compose_video_with_overlay: free clips plus marketplace stock, then weave. Use stock when the claim is generic (hands, city, kitchen). Generate I2V when the claim is this pack, this machine, this room. Mixing them in one cut is allowed if the palette matches. A generated bottle next to unrelated stock marble is two videos.

Do not generate a talking presenter and call it B-roll. A face is A-roll. AI UGC with a realistic person is a talking-head format. It inherits lipsync, likeness, and — if an EU audience would take the person as real — Article 50 marking at first exposure (in force 2 August 2026). Faceless product under a voice does not need that person.

Captions last, talking-head optional

Muted autoplay is still the audience. The TTS file is for whoever unmutes. The caption is for everyone else. Burn after the picture is locked to the voice. If you caption the VO against a black slate and then drop B-roll on top, type will sit on the pack label you just generated.

Add captions to a video transcribes the voice you actually laid. 165 language codes. 21 BASIC looks at 1×, 9 DYNAMIC at 2×. Caption styles is 45 CSS presets if you are in Caption Studio. Preview a few seconds of this B-roll, not a grey card. Keep the block in the safe area. Product plates often have lettering already; do not park karaoke on the SKU.

Spot-check names and numbers. ASR will miss the product name first. On a mute feed that misspelling is the whole message.

Talking-head is a fork, not a default:

No face. Voice + B-roll + captions. This page. Ship it.
Real face you already shot. Lay B-roll over parts of that A-roll to cover jump cuts and to illustrate. The voice may be the recording, or TTS if you are replacing the room sound. Do not generate a cousin of the person.
Synthetic presenter. That is a lipsync or avatar row, a different bill, a different disclosure. It is not required for a product read. If you add one, it is A-roll. B-roll still goes under the voice on the cuts where the face is not the point.

Do not lipsync a silent I2V pack shot. There is no mouth. Kling Lipsync is 6 credits on a video that already has a face; VEED Lipsync is 3 credits. Those are mouth passes. They do not turn a bottle into a host.

Native-audio video rows — Veo 3.1 at 64 credits for 4s (128 with sound on), 4 / 6 / 8s; Flux 3 Text to Video at 34 credits, 5–20s — bake sound into the generate. That is a different spine. If you use one, you are not “laying B-roll under TTS”; you are cutting a talking clip. You can still drop product I2V over parts of it. Caption after the mix is the mix you want transcribed.

The assembly order that does not double-bill

A usable order:

Script. One claim per beat.
TTS. 12 credits per 1,000 characters on Gemini 3.1 Flash TTS, or 4 on Cartesia Sonic 3.5. Lock the take.
Still(s). Approve the pack. Crop to 9:16 if the feed is vertical.
I2V (or stock B-roll) per beat. Silent rows. 6s / 10s on Hailuo; trim to the line.
Assemble: voice on A, pictures on V, cuts on the claims. Add B-roll to a video or compose_video_with_overlay / edit_video.
Optional face: only if the brief named one, and only as A-roll you can cut away from.
Captions on the locked mux. Burn for the feed. Export SRT only if a player asked.

Do not generate I2V first “to see what we have” and then write a script that matches the motion. You will narrate a twitch. Do not caption the TTS file against a colour bar. Do not run a native-audio hero and a TTS bed under it; you will get two voices.

Credits to watch, as catalog stickers: TTS is cheap next to video. A voice read at 12 credits per 1,000 characters plus three Hailuo Fast plates at 12 credits each is a faceless cut. Swapping in Hailuo Pro at 40 for a hero still is a resolution choice, not a voice choice. Both Hailuo cards are silent. A talking Veo take at 128 credits (4s with sound) is a different film. Pick the spine before you pick the row.

UGC-shaped ads still show about 4× click-through and about 50% lower CPC in the Kapwing 2026 / Billo compilations, and 86% of brands say UGC lifts performance. Those lifts do not require a generated face. They require a product in the picture and a line a mute scroller can read. Hands and a voice are enough.

If you need the studio to generate the voice, the plates, and the burn, sign up. Lay the read first. The pictures go under it. Captions go last.

FAQ

Do I need a talking head if I already have TTS and product footage?

No. A talking head is optional. Faceless UGC is B-roll under a voice. Add a face only when the brief names a presenter you can actually show.

Can I prompt the I2V model to speak the line?

Not on the silent plates this assembly uses. Hailuo 2.3 Fast has unset audio; Hailuo 2.3 Pro is audio false; Kling 3 Turbo is silent. Generate speech on a TTS row. Lay it. Caption it.

When should I use stock B-roll instead of generating I2V?

When the claim is generic and the SKU does not have to hold: hands, city, kitchen, reaction. Generate I2V when the picture has to be this product. Mix only if the look still reads as one cut.