Machine Translation Post-Editing: A Practical Workflow

What post-editing is, and how it differs from translating

A translator starts from the source and produces the target text. A post-editor starts from a machine draft and decides, line by line, whether to keep, correct or rewrite it. The skill is different: post-editors must resist rewriting acceptable sentences to their own taste, while catching the fluent-looking errors machine translation produces. The international standard ISO 18587 describes requirements for full post-editing of machine translation output, including the competences a post-editor needs.

Post-editing is worth doing when the machine draft is mostly right. When it is mostly wrong, editing takes as long as translating and produces a worse result, which is why the first step is always to judge the draft.

Light post-editing versus full post-editing

Goal: Light: accurate and understandable. Full: publishable, equivalent to good human translation.
Meaning errors: Light: all fixed. Full: all fixed.
Terminology: Light: fixed where wrong. Full: fixed and made consistent with a glossary.
Style and register: Light: left alone unless confusing. Full: natural, consistent formality, suited to the audience.
Typical use: Light: internal training, research, gist for decision-making. Full: marketing, public courses, anything carrying your brand.

Choosing the level is a business decision. An internal compliance briefing watched once by employees needs accuracy, not elegance. A product launch video in a new market needs full post-editing, because awkward phrasing undermines the message even when every fact is right.

Deciding how much editing a video needs

Before editing anything, run a short sample review. The spot-check method in machine translation quality evaluation takes a few minutes: sample the opening, the most technical passage and the most conversational one, and count minor, major and critical errors. That tells you whether the draft supports light editing, needs full editing, or should not be edited at all.

Consider the shelf life too. A video that will be watched for years by thousands of people justifies more editing time than one that will be viewed a few times this week.

A workflow for downloaded transcripts and subtitles

For a translated video, the editable artifacts are the translated transcript and the SRT or VTT subtitle files. A workflow that keeps edits fast and safe:

Download both transcripts (spoken language and target language) and the subtitle file in the format you need.
Read the source-language transcript against the audio first. Mark misheard words and names; these explain many translation errors and tell you which target lines to check.
Open the translated subtitles in a subtitle editor such as Subtitle Edit or Aegisub, or a plain text editor, alongside the source transcript. Match lines by timestamp.
Work through the triage order in the next section, editing text only and leaving the timestamp lines untouched.
Play the video with the edited subtitles at normal speed and check that every line can be read before it disappears.
Save as UTF-8 so accented letters and non-Latin scripts survive, and keep the original machine output for comparison.

For the differences between the two subtitle formats, see SRT vs VTT.

What to fix first: a triage order

Fix errors in order of damage, so that if time runs out, what remains is cosmetic:

Critical meaning errors: dropped negations, wrong numbers and units, reversed instructions, anything that could mislead.
Names and terms: people, products and specialist vocabulary, checked against a glossary as described in translating names and technical terms.
Omissions and additions: clauses that vanished or content that was never said.
Consistency: one translation per key term and one level of formality throughout.
Fluency and style: unnatural phrasing, literal idioms, wrong register for the audience. Light post-editing stops before this tier unless something is confusing.
Subtitle fit: shorten lines that run too fast, which matters most in languages where translation expands.

Worked example: a 12-minute training video into Portuguese

Suppose the spot check finds no critical errors, two major ones and a handful of minor ones, and the audience is internal staff. That points to light post-editing. The editor fixes a "15" that should be "50" (traced to the source transcript), standardizes the term for "approval workflow", and leaves stylistically flat but correct sentences alone. The dub for 12 minutes cost 600 credits (60 cents); the editing took far longer than the processing, which is normal and is where the budget should go.

When to re-run instead of editing

Sometimes the problem is upstream and line-by-line editing treats symptoms. Re-running the job with a better input is faster when:

The source transcript is wrong in many places because of music, noise or overlapping speakers. Export a cleaner audio mix, or trim noisy sections, and run again.
A long intro or outro in another language, or loud music, threw off the recognition of the opening; trimming it often fixes the rest.
You chose the wrong target language or voice, or the wrong file.

Re-running the same unchanged file is not a reliable fix for translation mistakes, because the engine receives the same text. Change the input, then re-run. The cost is easy to weigh: a 20-minute file costs 20 credits to transcribe again and 1,000 credits ($1.00) to dub again, usually less than the time spent editing a fundamentally flawed draft.

Pitfalls and limits of post-editing

Edits to downloaded text do not change audio that was already generated. If a dubbed line says the wrong number, correcting the subtitle file leaves the voice unchanged.
Over-editing wastes time and can introduce errors. If a sentence is correct and clear, leave it.
Changing timestamps by accident, or breaking the numbering in an SRT file, can stop players from loading it.
Fatigue sets in fast; long files are better edited in sessions, with critical checks done first.
On-screen text in the picture is outside the transcript and needs separate handling.

Post-editing mydubly output

mydubly gives you the files this workflow assumes. A full translation job returns the translated MP4, the translated audio, SRT and VTT subtitles in the target language, and transcripts in both languages; transcript mode returns timestamped text and subtitles in the spoken language, optionally translated, without a voice. The transcripts do not label speakers, so for interviews, note speaker changes as you edit.

Because the voice track is generated from the machine translation, decide before dubbing how much the spoken version must reflect corrections. A cheap way to preview whether translation quality is good enough to be worth dubbing is to run transcript mode first at 1 credit per minute; for subtitled-only releases, edited SRT or VTT files are the finished product. Start from the video translator for dubbed output or the subtitle generator for text-only work.

Next step

Pick a translated video, download its two transcripts and subtitle file, and do one timed pass using the triage order above. Note how many minutes of editing each minute of video took; that number, more than any metric, tells you whether light or full post-editing is sustainable for your content. Then translate the next video with the video translator and reuse your glossary from the first.

Frequently asked questions

How long does post-editing take compared with translating from scratch?

It depends on the quality of the draft and the level of editing. Light post-editing of a good draft is usually much faster than translating from scratch; full post-editing of a weak draft can take nearly as long. Time a sample before committing to a schedule.

Can I correct the translation and have the AI voice read my corrected version?

Editing the downloaded transcript or subtitles changes those files, not the voice track already produced from the machine translation. If spoken accuracy matters, review the translation before relying on the dub, and fix serious source problems by improving the input and re-running.

What tools can I use to edit SRT and VTT files?

Free subtitle editors such as Subtitle Edit and Aegisub show timing and text together and warn about reading speed. Any plain text editor works too, as long as you leave the numbering and timestamp lines intact and save as UTF-8.

Is post-edited machine translation as good as human translation?

Full post-editing aims for quality equivalent to human translation, and with a capable editor and a decent draft it can reach it. Light post-editing deliberately stops short: it is accurate but may read flatly. Match the level to the purpose of the video.

Should the post-editor watch the video or just read the text?

Watching matters. Speaker gender, who is being addressed, what 'this' refers to on screen and whether subtitles are readable at speed are all invisible in the text alone. At minimum, play the critical passages while editing.