Replacing speech in an existing video is fundamentally different from creating a fresh AI presenter. HeyGen’s Video Translation combines translation, synthetic re-voicing and optional lip synchronization. The best operating mode depends on whether the speaker is visible, how much movement the footage contains and whether a localized voice has been approved.
The documented three-mode split
HeyGen’s help documentation distinguishes Audio Only, Speed and Precision. Audio Only replaces translated speech without attempting mouth adjustment; its published site rate is four credits per minute. Speed adds lip sync and is listed at six credits per minute. Precision is documented at ten credits per minute and is aimed at more difficult visible-speaker footage: changing camera angles, profile faces, occlusions and multiple speakers. These are HeyGen’s website credits, not its separately priced API units. Verify the source page’s update date and actual account ledger before procurement.
The platform’s August 2026 language list separates audio-only and lip-sync support. A language appearing on a marketing page does not guarantee every engine, dialect and voice style in each region. A localization producer should check both the source and target language, write a pronunciation list for names and confirm whether the target audience expects formal, neutral or regional expressions.
Prepare the input
A clean source track improves translation; music, crowd noise and overlapping speakers create challenges. Trim unnecessary lead-ins, identify all speakers and save the original transcript. Avoid hard-coded captions because the translation operation does not automatically replace text already baked into picture. If the speaker’s face is mostly hidden or the video consists of slides, Audio Only may be the more economical and visually honest choice.
For an on-camera speaker, test a thirty-second excerpt containing the hardest section: a profile view, a rapid phrase, a camera cut and an emphasized word. Speed may be sufficient for consistent front-facing material. Precision is worth evaluating only if its more expensive processing solves a visible issue. Neither mode certifies the words are correct; a fluent-looking mouth is not translation QA.
Approval and delivery
Have a native-language reviewer check meaning, timing, proper names, numerical claims, sensitive terminology and speaker attribution. Reconcile the translated script against screen graphics and burned-in subtitles. Deliver a separate subtitle file wherever possible, alongside a captioned accessibility version. Preserve source language, selected mode, target language, approval date and audio-license/likeness consents in the manifest.
Do not use lip sync to fabricate statements by a real person or to hide a substantive alteration to a news interview. Localization should be disclosed when it matters to the audience’s interpretation, and consent should cover synthetic re-voicing and facial changes. For specialist speech synthesis, refer to AI 05 rather than treating video translation credits as a generic text-to-speech tariff.
Worked production example
Imagine localizing a safety video from English into Spanish and Japanese. If the speaker is off-camera and diagrams dominate, audio-only translation with translated captions may be clearer and cheaper than synthetic mouth motion. For an on-camera instruction, test the fastest lip-sync mode on a short front-facing passage, then inspect profile and occlusion moments before authorizing a more costly mode. Compare all numbers and warnings against the approved translation. Replace burned-in English signs in the edit rather than assuming speech dubbing will alter them automatically.
Relevant alternatives and trade-offs
For simple non-visible narration use audio-only localization and conventional subtitles; for a separately consented enterprise presenter examine Synthesia or a live shoot.