Everyone says "just add subtitles". Then the video gets 200 views in the new language and nobody can explain why.
I have been building a small video-translation tool, so I read a lot of localisation post-mortems. The pattern is consistent: people pick the wrong output for the audience, then blame the translation.
Here is the decision framework I would use, in the order that actually matters.
Subtitles and dubbing solve different problems
Subtitles assume the viewer reads the target language. They work for search indexing, muted viewing, technical content, and audiences that already read English reasonably well.
Dubbing assumes the viewer would rather listen. It works for training, children's content, ads, and markets where English literacy is low.
The failure mode is picking one for the other's audience. An HR safety video with subtitles only fails on a factory floor where half the team reads slowly in any language. A developer tutorial dubbed into Spanish can feel slower than subtitles, because developers are used to reading technical English.
If you cannot decide, produce both from one upload. That is cheaper than guessing twice.
The consistency trap
Most tools translate subtitles and dubbing in two separate passes. The result is a subtitle that says "dashboard" while the voice says "control panel". Viewers notice within ten seconds, and it makes the whole video feel machine-made.
The fix is architectural, not editorial: transcribe once, translate once, render both outputs from that single source. If a tool cannot explain whether it does this, assume it does not.
Length is a localisation problem, not a translation problem
French runs 15–20% longer than English. German compounds produce single words that do not fit a subtitle line. Japanese has no spaces, so line breaking follows grammar. Turkish carries meaning in suffixes, so a clause becomes a word and a word becomes a line.
Machine translation gets the words right and the timing wrong. That is why every workflow needs a five-minute human pass on line breaks — not on meaning, on fitting.
What machine output is actually good at
- First drafts of technical speech
- Terminology that appears in a glossary you supply
- Bulk work: twelve lessons, forty videos, one consistent voice
What it is not good at
- Jokes and wordplay
- Legal or medical phrasing
- Brand names that must stay untranslated
- Anything where a wrong name damages trust
Plan the human pass around those four, not around "reviewing everything". Reviewing everything is how localisation projects die.
A workflow that survives contact with a real project
- Glossary first. 20–50 terms: product names, feature names, acronyms. Every downstream error traces back to a missing glossary entry.
- One clip, not the whole series. Translate the clip that best represents the channel. The free tier on most tools covers 60 seconds, which is enough to judge transcription quality.
- Fix the glossary, not the clip. If lesson 1 mistranslated a term, correct the list before lesson 2. Otherwise you will be fixing the same error twelve times.
- Batch after the glossary is stable. Consistency across a course matters more than perfection in one lesson.
- Human pass on names, idioms, line breaks. Everything else can ship as machine output.
Cost expectations, without the marketing
Ignore "free forever" pages. Compare three numbers for your own clip:
- total cost of a 5-minute video in your target language, taxes included
- time to first usable draft
- cost of the second revision after a native speaker finds a wrong name
The third number is the one nobody advertises and everybody pays.
For reference, the tool I work on (LinguaReel) quotes per minute after upload and before payment, returns SRT plus a dubbed MP3 plus a subtitled MP4 from one transcript, and offers a 60-second trial with no account. Disclosure: I work on it. The framework above is independent of which tool you pick.
The one question to ask before any localisation project
Does the audience read the target language comfortably?
If yes, subtitles first — cheaper, searchable, toggleable.
If no, dub — and accept that the voice will not be the original.
Everything else is execution.
What is the first localisation decision you make — the language, or the format?
Top comments (1)
The dashboard versus control panel example is the one I see most in product videos too, and translating once fixes it. I'd add a glossary step before translation: a short list of UI terms with the exact label the app shows in each language. Then the voice and captions match the screen, not just each other. Otherwise the video says Settings while the localized app on screen says Preferences.