Whisper SRT to Word

What a Whisper SRT actually contains

OpenAI Whisper, whisper.cpp, faster-whisper, and the caption buttons built on them all emit the same family of file: numbered cues, a start and end clock, and a short stretch of recognized speech. That file is a good subtitle track. It is a poor document. Cue numbers, --> lines, and line breaks chosen for a video player are still there when you open the .srt in Word.

The model also leaves fingerprints that a human captioner usually would not. Silence and music often become a real-looking sentence. The same short phrase can repeat for several cues. Punctuation is uneven: some builds emit commas and question marks, others emit a stream of words with almost none. Names of people are guessed from sound, so “Jordan” may show up as a near-homophone three times in one interview. None of that is a timing bug. It is text you either delete or you publish.

A typical bad tail looks like this — invented here so you can recognize the shape, not copied from a model log:

41
00:18:12,000 --> 00:18:15,400
Thanks for watching.

42
00:18:15,400 --> 00:18:18,800
Thanks for watching.

43
00:18:18,800 --> 00:18:22,000
[Music]

Those three cues are not part of the interview. If you convert first and hope the formatter will notice, they become the last paragraphs of the .docx. Delete them in the SRT, or delete the paragraphs in Word. The converter will not classify a line as a hallucination.

Word-level files versus segment files

Some pipelines write one cue per word so a player can highlight tokens. Others write one cue per phrase, which is what you want for a document. If you still have the original audio and the tool that produced the SRT, re-export segments. If the word-level file is all you were sent, do not retype it.

Leave Merge same speaker on. Consecutive cues merge when the gap between them is about 2.5 seconds or less. Unlabeled Whisper cues count as the same speaker, so a run of one-word cues becomes a paragraph instead of a stack of single words. That merge joins text with spaces and collapses extra whitespace. It does not insert punctuation the model skipped, and it does not split a paragraph when the topic changes. A 40-minute unlabeled file can become a smaller set of long paragraphs. That is readable. It is not yet a script with turns.

If the word-level export also has wild timestamps — a cue that ends before it starts, or a gap of many seconds inside one sentence — fix the clocks before you rely on the merged paragraphs. Fix SRT timing is the checklist for overlaps and drift. Merging will not repair a clock that is offset by an hour.

A cleanup pass that is worth doing before download

Work down the file once. You are not proofreading every proper noun yet. You are removing text that should never reach a reader.

1. Search for stock hallucinations: “thanks for watching”, “subscribe”, “like and”, “[Music]”, “[Applause]”, “[BLANK_AUDIO]”, and a line that is only a period or a dash. Keep a genuine “[laughs]” if the transcript style your team uses wants non-speech. Delete the rest.
2. Collapse immediate repeats. If cue 12 and cue 13 are the same sentence because the model looped on room tone, delete one. Two identical paragraphs in Word look like a publishing error, not like emphasis.
3. Skim the first and last minute. Whisper often invents a greeting before speech starts and a sign-off after it ends. The middle of a clear recording is usually better than the edges.
4. If you know the speakers, add a prefix to the cue text: [Alex] or Alex: at the start of the line. The converter reads that prefix as a bold speaker label and merges only consecutive cues that share it. Without a prefix, everyone is one voice.
5. Save as UTF-8. A Windows notepad “ANSI” save will turn accented names into garbage before Word ever opens. The parser strips a UTF-8 BOM if one is present.

Do this in a text editor or in Subtitle Edit / Aegisub if you also need to nudge times. Then paste or drop the file on the SRT to Word converter. You do not need a second transcription pass unless the audio was unintelligible to begin with. Re-running Whisper and hoping for a cleaner paragraph is how people lose an afternoon.

Two Word files from one paste

For the pass where you still have the audio open, leave Show timestamps on. Each block in the .docx shows a short start–end range (hours appear only when the cue is past 59 minutes). When a sentence looks wrong, the range tells you where to listen. It does not seek the video for you.

For the copy a client, a journalist, or a newsletter will read, download again with timestamps off. Same paste, second file. Hide the clocks; do not delete the original .srt. The moment someone asks which minute a quote came from, the caption file is the index and the clean document is the prose. SRT without timestamps walks through that choice on its own.

Punctuation still belongs to a human pass inside Word. Find double spaces. Find a comma spliced onto the next cue because two segments were merged. Capitalize the sentence starts the model left lower-case if your style guide cares. That edit is faster in a real document than inside cue 300 of an SRT.

What not to expect from the .docx

• No diarization. If the SRT has no speaker marks, Word will not invent “Host” and “Guest”.
• No deletion of bad lines. Hallucinations, lyrics, and ads stay until you remove them.
• No word-level highlight track. Karaoke-style timing is flattened to paragraphs on purpose.
• No round trip back to SRT. If you still need a caption file for YouTube or a player, keep the cleaned .srt beside the .docx.

Interviews with more than one voice are a different cleanup than hallucination removal. Once the text is honest, label the turns and export a script from interview SRT to Word. If the source was a meeting recorder rather than Whisper, start from Zoom transcript to Word instead of forcing the VTT through a Whisper checklist.

Related workflows

• Fix SRT Timing Before Word
• Interview SRT to Word
• SRT Without Timestamps
• All subtitle-to-Word guides
• SRT to Word converter — same tool the buttons above open, with these toggles already set.