Dubbing vs. Subtitles vs. Native Generation for Localizing Ads

Search "translate video ad" and every result promises the same thing: your ad, in another language, in minutes. They don't do the same thing. Every video localization tool uses one of exactly three mechanisms: subtitle overlay, AI dubbing, or native generation. One translates what the viewer reads, one translates what the viewer hears, and one doesn't translate at all, it makes the video in the target language from the start. Once you know which mechanism a tool uses, you can predict how the result will feel to a viewer in that market before you spend anything.

The distinction matters most for paid creative. Organic viewers forgive a caption strip. Ad viewers scrolling TikTok or Reels in Jakarta or São Paulo decide almost instantly whether something was made for them or imported at them, and the mechanism decides which signal your ad sends. Taking one winning ad into Spanish, Portuguese or German comes down to two of them: dub it or remake it (compared below).

Mechanism one: subtitles

The oldest and cheapest option: keep the video exactly as it is and add a translated text strip. Every caption tool does this, and translation layers in CapCut, Veed, and Descript automate it well.

What it keeps: everything. The original performance, pacing, and audio are untouched.

What it costs you: the voice. On sound-on platforms the voice does the selling, and your viewer hears a language they may not speak while reading a translation of it. That is the register of imported content, documentaries, and airline safety videos, not of the native UGC that outperforms polished ads on TikTok. Subtitles localize comprehension, not persuasion.

Where it is the right call: organic content with an international tail, compliance requirements, and markets where viewers habitually watch foreign content subtitled.

Mechanism two: AI dubbing

The current mainstream of "video translation," and what most tools mean by localization. The pipeline is: transcribe the original speech, machine-translate the transcript, synthesize a new voice with text-to-speech, and warp the speaker's mouth so the lips roughly match. HeyGen's video translate, Synthesia's dubbing, Akool, and Topview's AI dubbing all work this way, and the large language counts in their marketing are counts of what their text-to-speech catalog can pronounce (the numbers change often, so check current details).

What it keeps: your footage. The same actor, the same scenes, now speaking Portuguese.

What it fakes: the performance and the intent. The voice is a synthetic track laid over a body that originally said something else, and lip-sync holds up on a front-facing talking head but degrades on cuts, side angles, and hands near the face. More fundamentally, dubbing translates sentences, but it cannot re-decide what the ad says. The hook that works in English ("I found this at Target...") gets faithfully translated into a sentence that means nothing in Indonesia. Prices stay in dollars. The humor stays American. Fluent audio, foreign ad.

Where it is the right call: you have finished footage that must be reused, a founder video, a testimonial, a piece of content whose exact footage is the point.

Mechanism three: native generation

The third mechanism skips translation entirely. Instead of adapting a finished video, the system writes the script in the target language first, with that language's own internet phrasing rather than translated English, then generates the video with the speaker performing that script in the scene. The voice isn't a text-to-speech layer over someone else's face; it is the performance the footage was generated around. Captions are aligned word-level to that voice and burned in with the right typography per script, Latin fonts for European languages, dedicated CJK fonts for Japanese and Mandarin.

This is how Riffkit localizes: it starts from a winning video, extracts the formula that makes it work (the hook structure, the pacing, the emotional beats), and regenerates that formula natively in the target language. The Brazilian version isn't the American ad wearing a Portuguese voice; it is a Brazilian ad built on the same proven structure. The same applies across all nine languages: English, Spanish, Portuguese, Indonesian, German, French, Italian, Japanese, and Mandarin.

What it keeps: the formula, the thing that actually made the original convert.

What it costs you: the original footage. Native generation makes new AI-generated creative; localizing the founder video you already shot is dubbing's job.

Which mechanism fits which job

• You must reuse finished footage (testimonial, founder story, event recap): AI dubbing. Accept that it will read as translated; pick front-facing footage to protect the lip-sync.
• You need comprehension, not persuasion (organic tail, compliance, docs): subtitles. Cheapest, most honest about what they are.
• You are producing new ad creative for another market: native generation. If the goal is an ad that a viewer in that market reads as local, translating a foreign ad, by text or by voice, starts one step behind an ad generated for them.

The pattern mirrors what we found mapping aspect-ratio tools: a crowded market that looks like thirty competing products is usually three mechanisms wearing different pricing pages. Ask which mechanism, and the choice mostly makes itself.

One winning ad, many markets: dub it or remake it

For paid creative, subtitles drop out early, so "translate a UGC ad into Spanish" or "run one video ad in five languages" comes down to two routes that answer different questions. Dubbing asks how to make this footage speak another language. A native remake asks what this ad would have been if it had been made in that market.

Dub when the actor is the asset. A founder, a creator your audience already trusts, a real customer's testimonial: the person is what sells, and a remake with a different person is no longer the same testimonial. If you hired the creator, check that your agreement covers a new market and a voice in a language they never recorded.

Remake natively when the ad has to feel made for the market, or the face isn't yours to reuse. If the hook or the joke only works in English, dubbing delivers it fluently and it still lands as an import. If the winning ad is a competitor's or a creator's TikTok you found, a dubbed version still runs their face; a remake keeps the formula, not the footage.

Dubbing leaves text filmed into the footage (a sign, the packaging) as it was; a remake writes its captions and post text in the new language. For an app, when the winning ad shows a flat screenshot fixed to the screen, choose On-camera placement and add a screenshot in that market's language as a product image: Adapt can show it in that layer as the exact pixels, laid over the finished video, so it stays readable. When the app appears only on a phone inside the scene, the video model draws that screen and its text can come out garbled, so keep the UI off the generated phone and let voiceover and captions carry the message (Off-camera placement).

How to make 3 language versions in Riffkit

1. Pick the winning ad once. A template, a TikTok link or an upload; a new link or upload becomes a template in your library, so the next versions start from it.
2. Choose Adapt. It writes new lines in the new language. Swap keeps the source's shots and language, so it's the mode for changing who or what is in a video.
3. Keep the same character and product. Pick your saved character so the same person appears in every version (keeping one character consistent), and the same product and placement: On-camera for something physical, or for an app when the winning ad shows a flat screenshot you want replaced with yours; Off-camera for other apps, sites and services.
4. Write the creative direction once. For an app with a screenshot per market, change only the screenshot's image name in each version's direction. Put any brand name or slogan that must read the same everywhere in double quotes: quoted text stays word for word, even inside a Japanese video. Everything else is written natively for each language.
5. Change only the language, one version at a time. Set Language to Spanish, check the price and submit; then Portuguese, then German. Each version is its own video.
6. Check each version's captions and post text. Caption fixes on a finished video need no new render, and extra vertical ratios (3:4, 1:1, 4:5) can be added to each version afterward; for a 16:9 version, set Aspect ratio to 16:9 when you submit.

On a free account the signup credit covers the first section of a video, enough to see the opening of the Spanish version before you decide on the rest; plans start at $15/month (Lite). Working from an agent? With the Riffkit skill in Claude Code, Codex or Cursor, or the connector in Claude or ChatGPT, ask for "this template in Spanish, Portuguese and German with my character": the agent quotes the price and asks before spending anything.

How to tell which mechanism a tool uses

Vendors rarely lead with the mechanism, but three questions expose it every time:

1. "Do I upload a finished video?" If the tool's input is your existing footage, it is subtitles or dubbing by definition; nothing that starts from finished footage can rewrite the script. If the input is a source video plus your product and a target language, you are looking at generation.
2. "Where does the translated text come from?" If there is a transcript-and-translate step anywhere in the flow, the script was written for another market and the output will carry that accent, however clean the audio. Native generation has no translate step to point to; the target-language script is the first artifact.
3. "What happens to the captions?" Dubbing pipelines re-time the original captions against a synthetic voice, and the drift shows on fast speech. Ask whether captions are aligned to the actual voice track word by word, and whether the tool renders non-Latin scripts with proper fonts and line-breaking. CJK line-wrapping is where caption pipelines quietly fail; a Japanese line has no spaces, and tools tuned on English overflow the frame.

Ten minutes with these three questions sorts any localization vendor list into its three mechanical piles.

The language-count trap

One buying note. Language counts in this category are text-to-speech catalog counts: if the vendor's speech engine can pronounce a language, it ships in the marketing number, which is how counts inflate into the hundreds once dialects are tallied separately. A catalog count tells you nothing about whether the script will be written in that language's native register, whether the captions will align to the voice, or whether the fonts will render the script correctly. Ask instead what the tool does per language, not how many it lists. A smaller number where every language is generated and validated end-to-end beats a large catalog of pronounceable ones.

For the market-by-market view of where each language earns its place, start with Spanish, the highest-leverage second language for US-facing sellers, and Brazilian Portuguese, the fastest-growing TikTok Shop market. Or skip the reading and riff a winner into another language directly.