How to check and edit SRT subtitles

Machine-generated subtitles are never 100% right, so proofread before you publish. SRT is plain text; once you know the format and the usual kinds of error, a 10-minute video does not take long. If what you want is a plain-text transcript without timecodes (video to text), there is a method at the end.

The SRT format is four lines

Each cue is a number, a timecode, the text, and a blank line:

1
00:00:00,060 --> 00:00:02,900
I want to have a chat with you today.
今天我想和大家聊一聊，

2
00:00:02,900 --> 00:00:05,300
How can a small team
spend very little money?
一个小团队怎样用很少的钱

• The timecode is written hh:mm:ss,mmm --> hh:mm:ss,mmm; note the comma before the milliseconds.
• A bilingual cue is simply a text part with two or more lines.
• There must be one blank line between cues.

What to edit with

• A text editor: for fixing a few typos, macOS TextEdit (saved as plain text) or any code editor is enough.
• An editing app: import the SRT, fix it against the picture, then export. Good if you also want to adjust the styling.
• A dedicated subtitle editor: for many cues or heavy timing changes, for example the open-source Aegisub.

Twinsubs only generates files for now and has no built-in editor. That is deliberate: proofreading is left to the tool you already know.

When you save, keep two things: the encoding stays UTF-8 and the extension stays .srt. If TextEdit saves as rich text by default, switch it to plain text first, or formatting data ends up in the file.

Five kinds of error in machine subtitles

1. Proper nounsNames of people, companies, products and terms. Recognition may write a similar-sounding word, and translation may render it literally. Check these first.
2. Numbers and unitsYears, amounts, percentages. One misheard digit changes the meaning, so check against the video.
3. Homophones and dropped phrasesWhere music covers the voice, half a sentence can be dropped or replaced by a similar-sounding word.
4. Break positionsA cue that ends mid-sentence reads awkwardly. Merge or split cues to fix it.
5. Overlapping or too-short timesTwo cues whose time ranges cross, or a cue shown for under a second so viewers cannot read it. Some editors refuse to import overlaps.

A proofreading order that saves time

1. Search proper nouns firstUse find and replace across the whole file to fix every wrong spelling of the same word at once.
2. Then scan the numbersLook only at cues that contain digits, and check them against the video.
3. Finally watch it throughLoad the subtitles in a player such as IINA, play from the start, and pause only where it reads badly.

A long video does not need to be listened to word by word. With this order, most errors are caught in the first two steps.

In a bilingual file, which line is wrong?

Each cue in a bilingual SRT has two lines: one in the target language and one in the source. When something reads badly, check the source line first:

• The source line is already wrong: recognition misheard. Fix the source line, then the matching target line.
• The source is right and the target is clumsy: a translation problem. Fix only the target line.
• Both lines are right but the meaning is broken: usually a break that cut one idea in two. Merge the two neighbouring cues, or move the boundary.

Turn an SRT into plain text (video to transcript)

What you want is not subtitles but a transcript with no numbers and no timecodes, for example for interview notes or course notes. Twinsubs writes only SRT for video, so the method is to make the SRT first and then strip the number lines and timecode lines. Pick either way:

Way one: a single Terminal command (tools built into macOS)

Choose the file first: for one language only, use the single-language file (such as Interview.en.srt); the bilingual file keeps both languages. Then run this in Terminal (replace the file name with yours, and quote it if it has spaces):

tr -d '\r' < Interview.en.srt | grep -vE '^[0-9]+$|-->' | sed '/^$/d' > Interview.txt

It does three things: strips Windows line endings, deletes the number-only lines and the lines with --> (the timecodes), and deletes blank lines. The result is saved as Interview.txt, and the original SRT is untouched.

Note: subtitle text that is only digits on a line (for example a line that says just “2024”) is deleted too, as if it were a cue number. Glance over the places where numbers appear after converting.

Way two: regex replace in a text editor

• Back up: Copy the SRT first and work on the copy.
• Delete the timecode lines: In an editor with regex support (VS Code, Sublime Text and so on), find ^\d\d:\d\d:\d\d,\d\d\d --> .*\n and replace it with nothing.
• Delete the number lines: Find ^\d+\n and replace it with nothing.
• Delete extra blank lines: Find \n{2,} and replace it with \n.

Each line of the result is one cue, not a paragraph. To turn it into running paragraphs, join neighbouring lines afterwards. If you need to know where in the video something was said, keep the SRT and do not convert.

Live translation has its own route: after a meeting you can choose to save a live session as a text file; see bilingual subtitles for meetings.

Subtitles are for reference; please proofread before publishing. See Terms of Use, section 3.

FAQ

Updated 2026-10-12. Back to the Twinsubs home page