An AI dubbing workflow replaces a video's narration with a translated voiceover made by text to speech. It runs in 6 steps: transcript, translate, adapt for timing, voice, sync and QA. This guide shows US and UK teams how to localize product, training and explainer videos into Spanish, French, German, Portuguese, Italian or Japanese.
Key Takeaways
- Dub when the audio delivers information, such as demos, training and explainers. Subtitle when the performance is the content.
- Timing is the hard part. Translated text usually runs longer than English, so test one segment per language before you voice the whole video.
- Fix timing in the words first, then with small speed changes. Re-edit the video only for flagship content.
- Kveeky voices your translated script in 40+ languages. It does not translate for you or match lip movements, so plan for a reviewed translation and voiceover-led footage.
- Never make a voice that sounds like a real person without their written consent. In the EU, AI audio that could pass as a real person must be disclosed.
Should you dub or subtitle?
Dub when the audio is a delivery mechanism. Subtitle when the audio is the content.
| Content type | Recommended | Why |
|---|---|---|
| Product demos and software walkthroughs | Dub | Viewers watch the screen, not the captions |
| Training and compliance | Dub | Reading captions while following a demo splits attention |
| Explainers and how-to videos | Dub | Same reason, and the narrator is rarely on camera |
| Marketing and brand films | It depends | If the original performance is the point, subtitle |
| Interviews and testimonials | Subtitle | A different voice speaking a real person’s words can mislead viewers |
| Drama, comedy and anything performed | Subtitle | AI voices can’t carry a full acting performance |
The AI dubbing workflow, step by step
Every localized video goes through the same 6 steps. Step 3 is the one most teams skip, and it’s where most rework comes from.
- Step 1: split the source into time-coded segmentsTranscript
- Step 2: translate for speech and get a native reviewTranslate
- Step 3: make each segment fit its time slotAdapt for timing
- Step 4: generate audio with one voice per languageVoice
- Step 5: place each clip at its timecodeSync
- Step 6: a native speaker watches the full videoQA
Choose a step for the details and the copy-ready templates.
Step 1: Build a time-coded segment sheet
Start from the final edit, not the original script. Narrators ad-lib, and editors cut lines.
- Split the narration into segments at natural pauses or scene changes. Give each one a start and end timecode.
- Separate spoken text from on-screen text. Titles, captions and UI labels are translated separately and have their own space limits.
- Remove idioms and local references before translation. "Hit the ground running" doesn't survive the trip.
- Mark do-not-translate terms, such as product names, feature names and UI labels that stay in English.
Segment sheet columns Paste into a spreadsheet
Step 2: Translate for the ear
A document translation and a voiceover script are different products. Ask for a translation meant to be spoken, with each segment's duration in view.
- Pick the market, not just the language. Spanish for Mexico or Spain, Portuguese for Brazil or Portugal, and French for France or Canada differ in vocabulary and accent.
- Use one glossary for every language, so your product doesn't end up with 3 names in one market.
- Set the register. Formal or informal "you" (tú or usted, tu or vous, du or Sie) is a brand decision. Make it once.
- Get a native-speaker review for anything customer-facing, even if a machine produced the first draft.
Translator brief Edit the brackets, then send
Step 3: Adapt each segment for timing
Translations usually come out longer than the English. A segment that runs long drifts out of sync with the action on screen.
Fix it in this order, from best result to most effort:
- Edit the wording to fit. Ask the translator to shorten long segments. This gives the most natural result.
- Adjust the speed a little. A small change is hard to hear. A large one sounds rushed, so go back to option 1.
- Re-edit the video. Extend a shot or hold a screen longer. Keep this for flagship content.
Timing check per segment Add to the segment sheet
Step 4: Generate the voice, one voice per language
- Choose one voice per language and keep it for the whole series. Record the voice name in your segment sheet.
- Listen for accent. Play the samples on the Spanish, French, German, Portuguese, Italian or Japanese page and pick a voice that suits your market.
- Match the original tone. If the English narrator is calm and measured, choose a similar voice and emotion preset.
- Add custom pronunciations for product and brand names so they sound the same in every language.
- Generate segment by segment and name each file with its segment ID. Long scripts are generated in parts anyway.
Download WAV for editing and MP3 for quick review rounds.
Step 5: Sync the audio to the video
- Work from the original project file if you have it. Mute only the narration track and keep the music and sound effects.
- Place each clip at its timecode in from the segment sheet, then nudge it to land on the matching action.
- Replace on-screen text with the translated titles and labels, and check they still fit their boxes.
- Add captions in the target language. Dubbed audio helps people who don't read English. Captions help viewers who are deaf or watching muted.
Generated audio does not follow mouth movements. Use it where the speaker isn't on camera: screen recordings, b-roll, animation and slides.
Step 6: QA with a native speaker
A native speaker of the target market should watch the full localized video with sound on, not just read the script.
- Log every issue against a segment ID and a timecode.
- Send wording problems back to step 3, and pronunciation problems to step 4.
- Use the full QA checklist further down this page before you publish.
Reviewer instructions Send with the video link
Have a reviewed translation? Paste one segment into Kveeky, pick a voice and check the timing.
Voice a test segment freeHow much longer will the translation be?
It depends on the language and the length of the text. W3C, the web standards body, publishes IBM’s average expansion rates for text translated from English into European languages:
| English source length | Average expansion |
|---|---|
| Up to 10 characters | 200–300% |
| 11–20 characters | 180–200% |
| 21–30 characters | 160–180% |
| 31–50 characters | 140–160% |
| 51–70 characters | 151–170% |
| Over 70 characters | 130% |
These figures describe written text, so use them for on-screen text: titles, buttons and lower thirds. Short labels expand the most, so leave room around them.
They don’t tell you how long a translated voiceover will run. Speaking pace differs by language and by voice, so measure it yourself:
- Pick a typical 30–60 second segment with normal pacing, not the intro or the call to action.
- Generate the English and the translated version with the voices you plan to use, both at normal speed.
- Divide the translated duration by the English duration. That's your timing ratio for this language and voice.
- Apply the ratio to the whole script to spot which segments will run long, and send those back for shortening before you voice the rest.
The data covers European languages only. For Japanese and other non-Latin scripts, character counts don’t predict spoken length, so the measured ratio is the number to trust.
What AI dubbing can’t do
Knowing the limits up front saves you from finding them in production.
- Lip-sync. Kveeky generates voice audio. It doesn’t change mouth movements, so talking-head footage will look out of sync.
- Translation. You supply the translated script. Kveeky voices it but doesn’t translate it for you.
- Acting. Emotion presets help with tone, but a full dramatic performance still needs a human voice actor.
- Cultural adaptation. An example that works in Ohio or Leeds may not land in Lyon or Osaka. Only a native reviewer catches that.
How Kveeky helps you localize video voiceovers
Voice your first translated segment free
Standard voices on the free plan. No credit card needed.
Consent and disclosure for AI voices
Use a voice that sounds like a real person only with that person’s written consent. This includes a voice meant to resemble your original presenter or a well-known narrator. Kveeky’s terms prohibit using the service to impersonate any person.
In the EU, the AI Act defines a “deep fake” to include AI-generated audio that resembles existing persons and would falsely appear authentic. Under Article 50(4), deployers that generate such content must disclose that it is artificially generated or manipulated.
A short disclosure in the video description is good practice everywhere:
Disclosure line For the video description or end card
For the rights side of publishing, read can you use AI voiceovers commercially?
What does AI dubbing cost?
The cost structure changes more than the price. Traditional dubbing repeats a voice artist, a studio session and a re-edit for every language. With AI voices, most of the work moves into translation, timing and review.
| Cost driver | What decides it | How to keep it down |
|---|---|---|
| Translation and native review | Your translator’s or agency’s rates, and the number of words | Lock the English script before translating; changes multiply by the number of languages |
| Voice generation | Kveeky credits are used per character of generated speech | Longer translations use more credits, so tighten the text in step 3 |
| Re-generations | How many segments fail timing or QA | Test one segment per language before voicing the whole video |
| Editing and sync | Whether you have the original project file with separate tracks | Ask for split narration, music and effects tracks when you commission the source video |
To estimate the voice part, add up the characters in each translated script, then compare the total with the plan minutes. The plan box below shows current plans, and the pricing page has full details.
For a worked example across several markets, see how to localize product videos for 5 markets.
QA checklist before you publish
| Check | Why it matters |
|---|---|
| A native speaker of the target market watched the full video with sound on | Reading the script misses timing, tone and pronunciation problems |
| Product names sound the same in every language | Custom pronunciations keep your brand consistent |
| Glossary terms are used consistently | Mixed terms make the product look unfinished |
| No segment starts before or ends after its action on screen | Audio that drifts makes demos hard to follow |
| Speed changes sound natural | Large speed changes sound rushed; edit the text instead |
| On-screen text is translated and fits its box | Expanded text often overflows titles and buttons |
| Captions in the target language are added and timed | Captions for prerecorded video are a Level A success criterion in WCAG 2.2 |
| Music and effects are at the same level as the original | A new voice track can sit too loud or too quiet in the mix |
| AI voice disclosure is in place where needed | Required in the EU for audio that could pass as a real person |
| Audio was generated on a paid plan for commercial use | Every paid Kveeky plan includes commercial usage rights |
If your localized video is training material, the training video voiceover guide and e-learning voiceover guide cover the production side. For demos, see product demo voiceovers. To fine-tune pacing in any language, use our guide to AI voice pacing, emphasis and breath. For accents and quality across languages, read multilingual voiceover with AI.
Frequently asked questions
What is an AI dubbing workflow?
It is the process of replacing a video’s narration with a translated AI voiceover. The usual steps are transcript, translation, timing adaptation, voice generation, sync and native-speaker QA.
Does Kveeky translate or lip-sync my video?
No. Kveeky turns your translated script into voice audio in 40+ languages. You supply a reviewed translation and sync the audio in your video editor.
How do I handle text expansion when dubbing?
Generate one translated segment and compare its length with the original. Shorten the wording first, then use small speed changes, because large ones sound rushed.
Do I need consent to use an AI voice that sounds like a real person?
Yes, get the person’s written consent first. In the EU, AI audio that resembles a real person and could seem authentic counts as a deep fake under the AI Act and must be disclosed.
Should I dub or subtitle my videos?
Dub product demos, training and explainers, where viewers watch the screen. Subtitle interviews, testimonials and performed content, and add same-language captions to dubbed videos.
Can I use AI-dubbed videos commercially?
Every paid Kveeky plan includes commercial usage rights for the audio you generate. Make sure you also hold the rights to the video, music and translation.
Sources
- Text expansion rates (IBM data, text translated from English into European languages): W3C Internationalization, “Text size in translation”, retrieved 9 October 2026.
- Deep fake definition: EU AI Act, Article 3(60), AI Act Service Desk (European Commission), retrieved 9 October 2026.
- Deep fake disclosure duty: EU AI Act, Article 50(4), AI Act Service Desk (European Commission), retrieved 9 October 2026.
- Captions for prerecorded video: W3C, Understanding WCAG 2.2 Success Criterion 1.2.2 Captions (Prerecorded), retrieved 9 October 2026.
- Credits per character and commercial usage rights: Kveeky pricing (machine-readable), retrieved 9 October 2026.
- No impersonation of any person: Kveeky Terms of Service, retrieved 9 October 2026.