SRT to text: preserve spoken numbers, not just clean formatting

A subtitle-to-text export can look clean while silently deleting part of the speech. The risky shortcut is removing every numeric line: a line containing 42 may be a spoken answer, not a cue identifier.

Treat the file as structured cues before deciding what to remove.

What a safe conversion preserves

This invented SRT example contains two cues:

1
00:00:01,000 --> 00:00:03,000
We reviewed 42 applications

2
00:00:03,000 --> 00:00:05,000
before the Friday meeting.

The reading copy becomes:

We reviewed 42 applications before the Friday meeting.

The cue identifiers and timing ranges disappear. The spoken number stays. The sentence boundary comes from the speech, rather than from the blank line between cues.

For a conversion pipeline, keep an untouched source file and make these checks explicit:

1. Recognize cue boundaries, sequence metadata, and timing lines before extracting text.
2. Preserve numeric text inside a cue's payload.
3. Keep speaker labels and meaningful sound descriptions.
4. Join fragments only when they belong to the same sentence and speaker turn.
5. Flag repeated passages for review rather than deleting them unconditionally.
6. Compare the opening, ending, names, and numbers against the recording.

An export from a caption editor is a useful starting point, but the result still needs inspection. A renamed .txt file has not undergone a structural conversion.

WebVTT needs its own handling

A file beginning with WEBVTT is not simply SRT with different punctuation. MDN documents cue identifiers, timing settings, and NOTE, STYLE, and REGION blocks. A timestamp-only replacement can leave those blocks in the transcript. Voice annotations may carry speaker information that should become a readable label.

Use a format-aware parser or export tool, then inspect examples with speaker changes and formatting. See MDN's WebVTT format reference.

Keep a path back to the recording

A reading copy can omit most timestamps. A reference copy can retain an occasional timestamp at a speaker turn or section boundary. Keep the original caption file either way: it makes an uncertain phrase much easier to locate.

W3C WAI also explains that turning captions into a descriptive transcript may require adding visual information. If a presenter says “choose this setting,” check the picture and label the editorial description separately from the spoken words. See WAI's transcript guidance.

Disclosure: I work on Wordtake. The full SRT-to-text workflow and review checklist covers source versions, reading copies, speaker changes, encoding checks, and publication. This is an editing workflow for existing subtitle files; it does not claim the product imports them.