A production company can use an AI video translator to create subtitle or dubbed drafts from a locked source video, then have a native speaker check meaning, timing, names and on-screen text. Choose dubbing only when it helps viewers follow the video; test lip sync on difficult shots. Cost each language by source minutes, processing, review and revisions before promising delivery.
The machine output is a localisation draft. It may translate the spoken words while leaving product screens, charts and captions unchanged. It may sound natural but choose the wrong term for a service. The client should approve a glossary and final wording, while a native-speaker reviewer watches the whole export rather than reading text alone. One approved language is a useful pilot before processing several.
Lock the source cut and collect editable assets
Begin with the final approved source video, its script, captions, music-and-effects track, voice track, editable graphics and brand terminology. If the source script is not exact, correct it from the audio first. Translation of a flawed transcript creates a second layer of errors. Give each language a version number tied to a single source cut, such as Source v3. A later source edit should trigger a change review for every language, even when it appears visually minor.
List the words that must stay consistent: product names, job titles, legal phrases, service levels, interface labels and the call to action. The client can supply approved translations or flag terms that must remain in the source language. Add pronunciation notes for names and acronyms. A glossary should include context; “plan” could mean a subscription, an architectural drawing or an intention. Without context, a fluent translation can still be wrong.
Ask for the destination of each version before making it. A training portal may support selectable captions and separate audio tracks; a social platform may need captions burned into the picture; a conference screen may have no reliable sound. The same translated script can produce different delivery assets. Record aspect ratio, caption format, maximum file size and whether the viewer can turn captions off. The production company should test the client's actual destination with a short file. A technically valid subtitle export is useless if the platform ignores it.
Take a six-minute staff onboarding video for an optician contains two speakers, eight on-screen labels, a diagram with three steps and a final sign-in instruction. The source has an editable caption file and separate voice and music tracks, but the diagram labels are baked into a flattened image. Before translation, the producer asks the client for the diagram source. Otherwise the translated voice may refer to steps whose visible labels remain unchanged. A source-script drafting process helps prevent this problem upstream.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowChoose between captions, voice replacement and lip sync
Subtitles preserve the original performance and are usually easier to correct: edit text, timing and line breaks without recreating a voice. They require viewers to read while watching the picture, so dense charts or fast demonstrations can compete for attention. A dub replaces or overlays speech in the target language, helping viewers follow visuals, but introduces voice quality, timing and rights questions. Lip sync goes further by altering visible mouth movement; it is only relevant when the speaker's face is visible enough for mismatched speech to distract.
Do not use lip sync by habit. A screen recording with a narrator needs translated narration and captions, not artificial mouth movement. A talking-head executive message may benefit from lip sync, but the translated sentence can be longer than the original and a side-profile shot can look awkward. HeyGen's Video Translation 101 help page offers three modes: Audio Only, with no lip sync; Speed, with lip sync tuned for front-facing video; and Precision, which it says copes better with side profiles, hands or microphones in front of the mouth and multi-person conversations. It also asks for speakers to face the camera at no more than 45 degrees and for one person to speak at a time. Test a representative section before committing to that treatment.
Consider accessibility alongside translation. A caption file should identify relevant speakers and meaningful sounds when the deliverable calls for accessible captions, while a translated subtitle may focus on speech. Decide which asset the client needs instead of calling every text track “captions”. Check whether the target platform supports separate tracks for each language and whether viewers can select them. If the project requires a single burned-in version, leave enough space around lower thirds and controls. A review on a large monitor alone will miss problems on a narrow screen.
For the illustrative onboarding video, the introduction has a front-facing speaker for 40 seconds, while the remaining five minutes use interface shots and diagrams. A sensible plan might translate the narration and captions throughout, then decide whether lip sync is worth applying only to the introduction. If the product cannot process separate sections conveniently, compare a full dub against a simpler subtitle version. The client should decide based on viewer needs and quality, not on the availability of a flashy feature.
Run a one-minute language pilot
Select a minute with the hardest material: a product term, a speaker change, a chart and the call to action. Translate that sample into one target language. Have a native speaker review accuracy, idiom, pronunciation, pace and whether the visible information still matches. The pilot should include the same resolution and aspect ratio as final delivery. A low-resolution preview can conceal caption crowding and lip-sync defects that become obvious in the export.
Provide the reviewer with the source script, glossary, client intent and a form for timestamped notes. Ask them to classify a change as meaning, terminology, pronunciation, timing, display or preference. “Sounds odd” is a useful starting signal but not enough for an editor to fix. A precise note might say: “02:14, translated voice uses a term for a retail purchase; the source refers to booking a clinical appointment; use approved glossary term.” This gives the production team a concrete correction.
Give the reviewer permission to stop approval when the source itself is ambiguous. If the original narration says “we process it within two days” but “it” could mean an enquiry or a payment, translation forces a choice the producer cannot safely make. Return the question to the client, repair the source script and decide whether the source audio also needs editing. That may be inconvenient late in production, but a multilingual release magnifies an unclear source sentence across every version. The review log should show that the issue began in the source, not in the translator.
Illustrative error: the source says “bring your current prescription” and the subtitle draft says “bring your current receipt”. Both words can fit the surrounding scene, but the instruction changes. The native reviewer flags the line, the client confirms the correct term and the producer updates captions and dub script together. If only the subtitle is corrected, the spoken track and text will conflict. Keep one approved target script as the source for all target-language outputs.
Review meaning before rendering the final voice
Use an editable translated script or proofread stage where available. HeyGen's help describes a Review & Edit (proofread) step on its Pro, Business, Team and Enterprise plans; it also says an uploaded SRT file is used as a reference and re-translated, so the output will not match that file word for word. That distinction matters when a client's approved phrase must be spoken exactly. Confirm the tool's current plan and workflow before promising exact phrasing. If the platform cannot lock the target words, use another route or get the client to approve the actual output after rendering.
Do two reading passes. First compare each target sentence with the source meaning and glossary. Second read the target alone to catch unnatural phrasing, grammar and tone. A direct translation can be correct yet unhelpful in a corporate training video. For example, an instruction intended to be concise may expand into a long, formal sentence that overruns a three-second graphic. The reviewer can suggest a shorter equivalent, but the client should confirm that no requirement has been lost.
Check numbers, dates, units and disclaimers manually. An AI output may translate “15” correctly but voice it ambiguously, or leave a unit untranslated in a chart. Do not let the reviewer silently change a product figure to fit local convention; return the question to the client. A localisation review is not permission to rewrite the company's policy. Keep a decision log for any meaning change so later language variants can be treated consistently.
Inspect the video, not just the transcript
Watch the exported version from start to finish at normal speed. Then inspect every flagged timestamp frame by frame where needed. Check captions for reading time, line breaks, safe placement and overlap with lower thirds or interface controls. Check that every caption belongs to the right speaker and remains visible long enough to read. The subtitle accuracy and delivery checks can support that separate pass.
For dubbed footage, listen with headphones and ordinary speakers. Check whether music masks the new voice, whether the original voice leaks through, whether pauses match actions and whether one voice changes character midway. With lip sync, examine visible faces at cuts, hands across mouths and side angles. An excellent translated sentence can still be unusable when a mouth continues moving after the sound ends. Choose a cutaway, shorten a line with native-speaker approval or use a plain dub without lip sync.
Timing failures look like this: a target-language phrase is 2.5 seconds longer than the source, so the voice explains step three while the diagram still shows step two. The editor can extend the diagram, move the narration or split the line. Simply raising playback speed makes the voice unnatural and may not fix the instructional mismatch. This is why the final reviewer needs the video, not only an exported subtitle file.
Use a second illustrative pass on a data-heavy chart. Suppose the source voice says “sales rose from 12 to 18 units” while the graphic labels show 12 and 18. The target dub is correct, but automatic captions display 12 and 80 because of a speech recognition error. A reviewer listening without looking may approve the voice; a reviewer reading without listening may blame translation. The final check compares source evidence, target audio, captions and graphic at the same timestamp. Any number that appears in more than one medium should agree across all of them.
Inspect on-screen text separately. HeyGen's translation help explicitly says text baked into the video image is not translated by its video translation feature. For the optician video, the eight labels and three diagram steps need editable replacements or an approved overlay plan. Check font support, line length and contrast after substitution. A translated label can be longer and collide with an icon. The client should approve final screen text as well as narration.
Calculate a per-minute quote that includes review
Do not compare platforms by a vague “per minute” figure. Identify whether the charge applies to source duration, each target language, reruns, proofread steps or final output. As of November 2026, HeyGen's API pricing page lists video translation at $0.57 a minute for audio only, $0.81 for Speed lip sync and $1.50 for Precision lip sync, charged by the actual seconds generated. Its web app deducts plan credits instead: 4 a minute for Audio Only, 6 for Speed and 10 for Precision. Those are separate purchasing routes. Happy Scribe's help says translating a 30-minute transcript consumes transcription minutes plus translation minutes for each language. Confirm the current plan, minimum charge and rounding rule in the actual account before quoting a client.
Use a sheet with source minutes, language count, quoted processing rate, expected reruns, native-speaker review minutes, edit minutes and project management. Run the sum for a six-minute video in three languages on HeyGen's $0.81 API rate for Speed lip sync: 18 output minutes cost $14.58 for the first pass. If one language needs a full rerun, processing becomes $19.44. At an illustrative 45 reviewer minutes and 30 editor minutes per language, the human work is 225 minutes. The $19.44 processing line is therefore a small part of the job; price the human work using the studio's own rates.
A second illustrative scenario is a 30-second social clip in one language. HeyGen's API bills by the second, so 30 seconds of audio-only translation comes to about $0.29, but another platform may charge a minimum unit or round each file up to a full minute; check before assuming a half-minute charge. The producer still needs to open the file, check the call to action, inspect captions and export the required aspect ratios. A shorter clip can have a higher cost per finished minute because the fixed review and delivery tasks remain.
Build a sensitivity range into the quote. For a six-minute source and three languages, ask what happens if each target needs one complete rerender, not only one of them. At the $0.81 Speed rate, three extra six-minute passes add $14.58, while the more significant cost may be another native-speaker watch-through and edit. Decide whether the quote includes one correction round, how many target-language words the client may change, and who pays for a changed source cut. Put those assumptions in the proposal. Without them, a low apparent per-minute rate hides an open-ended revision obligation.
Separate revisions caused by translation mistakes from revisions caused by a client changing the source. If the client changes a product claim after approving Source v3, every language may need a new voice render, caption pass and screen-text edit. Quote that as a change to the source, not a free correction to the translator. In the revision log, name the source version, target language, timestamp, reason and person who approved the new line. That record stops old language variants from being accidentally republished.
Know when a human voice is the right production choice
A human dubbing artist is worth considering when performance carries the message: a personal apology, testimony, executive reassurance, humour or a sensitive announcement. A director can adjust pace and emphasis in response to the picture. Human casting also offers a clear route to approved pronunciation and brand tone. It still requires a translated script, recording, edit and review; it is not automatically accurate because a person speaks it. The relevant comparison is the finished deliverable and the rights attached to the voice.
Check consent and licence terms before reproducing a real person's voice with AI. The fact that the source video contains a speaker does not automatically authorise synthetic reuse in another language. Ask the client who has the rights to the source performance, which languages and channels are covered, and whether approval is needed from the speaker. If the answer is uncertain, use subtitles, a licensed replacement voice or a human performer under an appropriate agreement. Voice-over consent deserves its own documented decision.
For a routine internal software demonstration, an AI dub with native-speaker checking may be suitable. For a founder's personal account of a serious incident, a neutral subtitle or human performance may preserve trust better than a synthetic imitation. This is a production judgement, not a technical score. Show the client the one-minute pilot, explain visible trade-offs and record the selected route before making every language version.
Deliver a language package that can be maintained
For each target language, deliver the approved video, separate captions where requested, target script, glossary, version reference and a list of unresolved limitations. Name files so the language and source cut are unmistakable. Keep editable project assets and approval notes under the agreed retention policy. If the client later corrects one phrase, the producer should be able to locate it without re-translating the entire film from a flattened export.
Use a final checklist: correct source version; native-speaker sign-off; glossary followed; claims and numbers approved; captions checked on devices; on-screen text translated; dub and music balanced; lip sync inspected where used; voice rights confirmed; and client approval recorded. One person should own the final pass across all language versions. Translation software can generate variants quickly, which makes version control more important rather than less.
After delivery, archive the test minute and one approved full export for future comparison under the client's retention agreement. If a new product update changes the voice or translation engine, rerun the same difficult section before using it on a new campaign. Record defects by type rather than one general “quality” score: mistranslation, subtitle timing, unsupported claim, audio mix, graphic mismatch and lip-sync artefact. That record tells the producer whether a new tool solves the actual bottleneck or merely makes the first render faster.
Questions before a localised video goes live
Should a corporate video use subtitles or AI dubbing?
Choose subtitles when the original voice matters, the audience accepts reading and the budget or review time is limited. Consider dubbing when viewers need to follow visuals without reading, but add native-speaker review of wording and delivery. Test both on one representative minute before ordering every language.
Will lip sync translate text shown in the video?
No. Lip sync concerns visible mouth movement. Titles, charts, captions and graphics need their own localisation check. HeyGen's help page explicitly says its video translation does not translate text baked into the image, so production teams should keep editable source graphics.
When is a human dubbing artist the better choice?
Use a human performer and director when nuanced acting, sensitive testimony, strict brand voice, complex dialogue or contractual voice rights make an automated dub hard to approve. AI can still support a rough translation, but the client should approve the final script and the performer arrangement.
Further reads
- Subtitling Rates per Minute: What to Charge When AI Drafts First — Understand human review and subtitle pricing.
- AI Video Production Software Compared for a Small Studio — Compare the wider editing stack.
- AI Shot List Generator: Build a Shoot-Day Plan From the Script — Plan shots that leave room for captions.
- Best AI Filmmaking Tools by Job for a Small Production Business — Place localisation in a wider tool workflow.
- Video Production Proposal: Scope, Budget and Timeline Drafted With AI — Quote localisation as explicit project scope.
- AI Video Content Agency: Services a Small Studio Can Add and Price — A six-service menu for a small video studio, a worked pricing floor, and the rights, consent and disclosure checks before delivery.
- Descript vs Riverside for Podcast Producers: Which to Standardise On — How to choose between Descript and Riverside for client shows: a timed test, plan limits, multi-client set-up, AI features and data terms.
- AI Subtitle Translation for Agencies: Quality Checks Before Delivery — A cue-by-cue workflow for AI subtitle translation, with timing, line-length and meaning checks, a QC sample and a brief for the native reviewer.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: HeyGen Help, Video Translation 101; HeyGen Help, HeyGen API Pricing Explained; Happy Scribe Help, Step-by-step: How to translate your subtitles; Happy Scribe Help, How can I translate my transcript from the Editor? Prices checked November 2026.